50 comments

  • jeanlucas2 hours ago
    I was wondering why my local tool to inspect coding agent sessions stopped working in some cases.<p>This is a really interesting engineering decision, I wonder how many people will want an encrypted external piece of instructions running on their machine.
    • EGreg2 hours ago
      It seems their incentives aren’t exactly aligned with their users, including corporate users. Look at the latest statements from Alex Karp, and now Satya Nadella etc.<p>Is this user-hostile? Encrypting stuff from the user is what RIAA used to do with DRM, worried about copyright infringement.
      • pornel1 hour ago
        Nothing by Alex Karp is aligned with any users. This guy is a mercenary, and he seems to be in this job for the love of killing.
        • ctkhn1 hour ago
          I thought tweaking was his highest calling
      • ssgodderidge1 hour ago
        [dead]
    • baby1 hour ago
      if you&#x27;re running in YOLO mode you already don&#x27;t care, what matters are the tool calls and these can&#x27;t be encrypted
      • Majromax49 minutes ago
        No, you&#x27;d still care. YOLO mode is about instantaneous permissions and access control, inspection of subagent prompts is about retrospective <i>quality</i> control. If the main model is instructing subagents to do a subtly wrong thing, the overall process quality will degrade in ways that might be very hard to detect or fix without deep inspection of the middle stages.
      • embedding-shape57 minutes ago
        I&#x27;m running in YOLO mode, not sure why that mean I don&#x27;t want introspection into what the sub-agents are running? Everything is running in a constrained environment, so not sure why you&#x27;d need the harness to have limits.
    • echelon1 hour ago
      [flagged]
      • bronson1 hour ago
        You can! Subscribe to a Z.ai coding plan.
    • fidotron1 hour ago
      [flagged]
      • b40d-48b2-979e1 hour ago
        <p><pre><code> training people to believe that them having impossible to inspect encrypted channels from your machines to theirs is a best practice </code></pre> Oh yeah, I much preferred when my government and whoever controls the infrastructure could snoop on my traffic instead!
        • airstrike1 hour ago
          idk about you but I&#x27;d rather not be murdered regardless of who the perpetrator might be
          • b40d-48b2-979e1 hour ago
            That is not the argument being made and this interpretation feels entirely in bad faith akin to the other sibling comments.
            • bobthebob1 hour ago
              Maybe to you, it resonates with me.
          • mrexroad1 hour ago
            Better make sure the pizza gets delivered on time, else you’ll have to pull your sword.
        • LeBit1 hour ago
          How did they do that?
          • cobertos1 hour ago
            Packet capture and analysis. It&#x27;s not that complicated
          • b40d-48b2-979e1 hour ago
            Does the word &quot;plaintext&quot; mean anything to you?
        • skywhopper1 hour ago
          You’ve been misled about how encryption works apparently.
  • niam4 hours ago
    This title is easy to misinterpret. If I understand correctly: Codex now encrypts <i>sub-agent</i> prompts and hides those prompts from the user.<p>edit: originally was &quot;Codex starts encrypting prompts, uses cyphertext for inference instead&quot;
    • hansihe2 hours ago
      It seems likely to me this was driven by the `ultra` mode in 5.6, which fans subagents to do work. This mode was previously only available in the web UI (what was previously known as pro?)<p>It seems possible they trained this by doing full RL rollouts of agents interacting with each other. They likely view these prompts somewhat the same as raw reasoning traces, they don&#x27;t want people to train directly on them.<p>I am unsure if this has been confirmed, but there are some signs that the opaque &quot;compaction blob&quot; they return from their dedicated compaction endpoint might not be text at all, rather a latent space representation of the conversation. The fact that OpenAIs compaction seems to be much higher fidelity than a lot of other providers makes me inclined to believe this.<p>If this is true, it doesn&#x27;t seem far fetched to infer that they might be applying similar techniques to prompting subagents.<p>I would be curious to see if this way of spawning subagents (encrypted blob) is used when subagents of a different model type is spawned.
      • tpurves2 hours ago
        &quot;Latent space representation&quot; I have been waiting for this moment in the evolution of AI. Well, waiting with some trepidation. It seems inevitable that frontier AI&#x27;s will, at some point, leave behind human-comprehensible representations of language. Purely for functional reasons, it&#x27;s going to start making sense for AI agents to communicate amongst themselves in much more efficient ways than borrowing the languages of flesh-bag humans as an interface medium.<p>I Imagine next that programming languages, interfaces and API design starts going this direction next. Being written, expressed and optimized as blobs of high dimensional vector space. As humans we might still be able to understand some abstractions of what our AI&#x27;s are talking about to each other, but maybe not more so then we understand how different regions of our own brain communicate with each other.
        • mparramon1 hour ago
          It has already started:<p><pre><code> zero-graph v1 origin source-text module &quot;hello&quot; hash &quot;graph:a7f7e6899a73f3b4&quot; node #decl_ad8d9028 Function name:&quot;main&quot; type:&quot;Void&quot; public:true fallible:true node #param_4610ae76 Param name:&quot;world&quot; type:&quot;World&quot; node #expr_c403020c MethodCall name:&quot;write&quot; type:&quot;Void&quot; node #expr_653eeb6e Literal type:&quot;String&quot; value:&quot;hello from zero\n&quot; edge #expr_c403020c arg #expr_653eeb6e order:0 </code></pre> <a href="https:&#x2F;&#x2F;zerolang.ai&#x2F;" rel="nofollow">https:&#x2F;&#x2F;zerolang.ai&#x2F;</a>
        • karma_daemon1 hour ago
          I&#x27;ve been thinking the same! Haha<p>Even for like token efficiency it could make sense - like imagine if the representation were more compact<p>Agents acn already translate languages quite well. It doesn&#x27;t seem crazy that they could work and think in a model specific language, and then translate back to English or something for the user
        • EGreg2 hours ago
          I strongly believe that the future is the other way. New programming languages and environments designed for strong auditability and preventing bugs will dominate. Only bad actors will use latent space representation, and it might even be outlawed. But the bad actors will proliferate underground…<p>I actually built such a language: <a href="https:&#x2F;&#x2F;magarshak.com&#x2F;U.html" rel="nofollow">https:&#x2F;&#x2F;magarshak.com&#x2F;U.html</a>
          • mparramon1 hour ago
            Every time a state tries to regulate, it loses to the states that promote instead. Seems to be a better evolutionary strategy.
          • michaellee81 hour ago
            is that called rust? that is the only thing i feel safe to let agents vibe code
        • TowerTall2 hours ago
          Google&#x27;s AI just created its own universal &#x27;language&#x27; (2016)<p><a href="https:&#x2F;&#x2F;www.wired.com&#x2F;story&#x2F;google-ai-language-create&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.wired.com&#x2F;story&#x2F;google-ai-language-create&#x2F;</a>
          • StilesCrisis40 minutes ago
            That&#x27;s not really a language. It was just a reporter dumbing down the idea of &quot;vectors in latent space&quot; which eventually became LLMs.
        • UltraSane59 minutes ago
          It seems like the most efficient method would be for LLMs to communicated by exchanged latent space representations directly. Serial language is a incredibly inefficient way to encode these, a lot like flattening a complex graph into text.
      • a-dub56 minutes ago
        &gt; It seems possible they trained this by doing full RL rollouts of agents interacting with each other. They likely view these prompts somewhat the same as raw reasoning traces, they don&#x27;t want people to train directly on them.<p>this tracks. anthropic protects these as well iirc.<p>&gt; I am unsure if this has been confirmed, but there are some signs that the opaque &quot;compaction blob&quot; they return from their dedicated compaction endpoint might not be text at all, rather a latent space representation of the conversation.<p>probably not a latent (to my knowledge latents aren&#x27;t really part of the outer loop in ar-transformer inference processes), but maybe non-human-readable reasoning traces as occurs in fable.
      • wren69912 hours ago
        I think you hit the nail on the head here. Having subagent dispatch in the loop for RLVR is something we&#x27;ve already seen in open models, like Kimi K2.5 and later, so it&#x27;s no great stretch to assume OpenAI are doing it too.<p>If you keep RL&#x27;ing the dispatch then the prompts are likely to keep diverging from the type of prompt a person would write (like CoT becoming increasingly incomprehensible), and that divergence is part of their competitive advantage.<p>&gt; rather a latent space representation of the conversation<p>Student&#x2F;teacher models derived from the same checkpoint convey a lot of latent information through token choice, as in: <a href="https:&#x2F;&#x2F;techxplore.com&#x2F;news&#x2F;2026-04-ai-chatbot-student-owls.html" rel="nofollow">https:&#x2F;&#x2F;techxplore.com&#x2F;news&#x2F;2026-04-ai-chatbot-student-owls....</a><p>I wonder if this is something they can take advantage of by training on compaction inside of the RLVR loop?
      • dannyw1 hour ago
        If there is no visible prompt at all, then that is very understandable. The PR issue exposes a real gap though: subagent spawns need a human-readable audit trial, of its goals&#x2F;intent, its boundaries and scope and limitations, etc; for basic responsible agentic harness functionality.<p>Hopefully they can add that.
        • embedding-shape1 hour ago
          &gt; Hopefully they can add that.<p>Add? Just make the sub-agents input prompt not encrypted, literally change &quot;encrypted: true&quot; to &quot;encrypted: false&quot; everywhere and everything continues to work as it used to.<p>They need to fix the regression, not add something new here.
      • dist-epoch1 hour ago
        &gt; latent space representation of the conversation<p>and how would you load that back into the model? they are token-in, token-out, plus the KV-cache which is derived from token-in
        • hansihe58 minutes ago
          They are not really token-in token-out per se, they are embedding-in embedding-out.<p>When operating on text, you embed each token into the LLMs embedding space. You go from a discrete token to a point in embedding space.<p>Likewise, when processing images, you have a image embedding model which produces a set of embedding vectors representing the contents of the image in the LLMs embedding (latent) space.<p>This same concept can be extended to compaction. Instead if limiting yourself to discrete tokens, you could generate a set of embedding vectors which represent the contents of the compacted conversation in latent space.<p>These have the possibility of containing a lot more semantic information per vector, which is why this can be appealing.<p>A big downside is decreased interpretability. AI safety people are generally fairly opposed to latent space reasoning for example, it can be harder to tell what the model is actually doing and if it is trying to deceive you.
    • themgt2 hours ago
      It&#x27;s sort of insane though, you not only have dozens&#x2F;hundreds of stochastic agents running on your machine, but you cannot even inspect the instructions those agents are working off of?<p>I&#x27;ve gone in to look at Claude subagent&#x2F;workflows and sometimes been like &quot;no this was a mistake to spin up&quot; ... Codex users just get to token yolo the encrypted telephone operator instructions+shell from orchestrator to subagents?
      • djeastm2 hours ago
        &gt;but you cannot even inspect the instructions those agents are working off of?<p>It makes more sense when you realize they don&#x27;t want developers to be doing any coding at all. That&#x27;s what they seem to be moving towards. From product manager to product via AI.
        • tempaccount4201 hour ago
          Last stage is moving everyone to their cloud platforms, they deploy everything for you, you don&#x27;t even get to see the code, just the deployed end product.<p>Because letting you look at the code would be too dangerous, you could reverse engineer an exploit to another product! Or distill their internet-distilled model!<p>But don&#x27;t worry, at least it will be very convenient.
      • Jean-Papoulos2 hours ago
        You already have an agent freely doing stuff on your machine. Subagents prompts are a weird place to draw a line. It&#x27;s not like you&#x27;re reading everything the agent is doing in any case, let&#x27;s not kid ourselves.
        • embedding-shape2 hours ago
          When things go wrong I very much read the session traces to figure out what in my prompt wasn&#x27;t good&#x2F;explicit enough, then retry to evaluate if it would have helped.<p>I was about to do the same with Sol + Ultra, but then discovered this encryption issue that prevents me from doing the same for sub-agents.
        • realusername1 hour ago
          &gt; It&#x27;s not like you&#x27;re reading everything the agent is doing in any case<p>Personally I do, these tools aren&#x27;t mature enough to be used without supervision
    • bartread3 hours ago
      I imagine this will be because a decent chunk of the IP in Codex is probably within its prompts, how they&#x27;re built, and how they&#x27;re sequenced and orchestrated, rather than in the codebase per se.<p>We had this discussion a few months ago where we talked about allowing people to choose an AI provider and provide their API key, thinking about enterprises with &quot;preferred&quot; (read: mandated) AI suppliers. We also wanted to offer the kind of very simple pricing that this is one way of enabling. But we realised pretty quickly that this would&#x2F;could lead to leaking our back end prompts to customers and, although those prompts are only a part of the value add, if you could build a detailed trace of them then you&#x27;d be able to relatively easily reverse engineer a lot of what we&#x27;re doing.<p>So we quickly dropped that idea.
      • agumonkey3 hours ago
        I&#x27;m unable to understand how much value can be in low-definability non deterministic prompts. It feels like the kept the right divinity spell into a chest.
        • hnlmorg3 hours ago
          I don’t disagree with your divinity spell comparison but unfortunately there is a lot of value in the prompts because these spells are the “programming languages” of LLMs.
          • agumonkey2 hours ago
            yeah i get it too, i&#x27;m just flabbergasted that this is today&#x27;s market<p>it reminds me of the pre-vulkan game programming days.. drivers were black boxes, game developpers had to resort to magic tricks to do stuff, until everybody got fed up and wanted some logical ground to operate
            • bartread2 hours ago
              It&#x27;s a brave new world, etc., isn&#x27;t it?<p>One does find oneself slightly askance at one&#x27;s own thinking sometimes, that&#x27;s for sure.<p>But I suppose, is it really so different? I mean, back in the day moreso than now, a lot of the valuable IP in any system was in the design and specification of that system - the problems usually solved within the design and specificaion (use X algorithm, etc.) - and the code was &quot;just&quot; the implementation of those solutions.<p>So perhaps it&#x27;s more of a regression in some ways: the value is in the specification (the prompt) once again.<p>Your point about stochastic behaviour is well made though, and there is no way to 100% guarantee or formally verify the behaviour of a system that relies on an underlying technology whose behaviour is fundamentally stochastic.
            • oblio2 hours ago
              Further proof that this tech stack is immature and would have needed to bake for a more years.<p>In an ideal world this would have been public tech like ARPANET or WWW and there would have been 2-3 major iterations (until the equivalent of Claude 7-8) and only then would everyone have tried to build huge businesses on top of it.<p>I mean, sure, it&#x27;s sort of usable, but the churn is insane. And we&#x27;re burning the planet (and probably the economy, too) for it.
        • embedding-shape2 hours ago
          &gt; understand how much value can be in low-definability non deterministic prompts<p>When was the last time you used an LLM to evaluate how true those last part(s) still are?<p>I also love how you went from &quot;I&#x27;m unable to understand&quot; to &quot;This is surely right&quot;, it&#x27;s a good representation of the software ecosystem at large :)
      • saidnooneever3 hours ago
        the trick about agentic systems is definitely how to do the prompting. things like automation and sandboxing are trivial in comparisson. if you generally ask via API model directly you can see what basic answers it actually yields and how fine tuning prompts and refinements to output as well as adversarial prompts etc are important to get relatively solid results.<p>a lot of expertise of certain domains&#x27; workflows is needed to make it functional within that domain. some of this can be yielded via prompting too etc so its also baoance of how much to prompt it vs. how much of it you wanna let it reason over itself. (if you tell it too much i lock it into a path and if you tell too little it will give incomplete results )
      • dmurray3 hours ago
        Perhaps AI providers should support this natively: the customer supplies the API key but doesn&#x27;t get access to the transcripts.
        • bartread2 hours ago
          I don&#x27;t know how you&#x27;d enforce that unless it was something you could mandate at the level of the API call, and then the API call is rejected if the customer hasn&#x27;t configured it for &quot;no transcript&quot;.<p>It sort of feels like an area of friction even still.
          • dmurray1 hour ago
            I was thinking of an API key that was scoped both to a specific customer and a specific service provider (perhaps both have to do something to provision it).<p>Billing goes to the customer, debug logs etc go to the service provider.
    • imhoguy3 hours ago
      Yeah I thought &quot;wow, some homomorphic encryption* stuff&quot;, but then &quot;nah, usual greed&quot;.<p>* <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Homomorphic_encryption" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Homomorphic_encryption</a>
      • embedding-shape3 hours ago
        The title was fixed like 40 minutes ago, when you come back to old browser tabs you probably want to hit that reload button before leaving a comment ;)
        • ofjcihen54 minutes ago
          Unnecessarily snide remark for someone just commenting on their interpretation.
          • embedding-shape50 minutes ago
            At that point 50% of all comments were about the title and it had been updated almost a whole hour before parent made their comment. Sorry for being low on patience.
    • postalcoder3 hours ago
      It&#x27;s also not the first time Codex started encrypting stuff. Their excellent compaction endpoint has served up a giant encrypted blob since at least five months ago.
      • hyperbovine2 hours ago
        Feels strongly like we&#x27;re in the late-stage-AI-unicorn phase. If this is really their moat then the Chinese companies will win.
        • realusername1 hour ago
          &gt; If this is really their moat then the Chinese companies will win.<p>I already switched to a Chinese provider personally, I don&#x27;t think the difference between both really justifies the wide gap in pricing
          • embedding-shape1 hour ago
            &gt; I don&#x27;t think the difference between both really justifies the wide gap in pricing<p>I&#x27;d be ecstatic if this was true, but nothing so far comes close to the SOTA models from OpenAI + highest reasoning, but I&#x27;d be more than happy to be proven wrong by testing it out myself.<p>So far, I&#x27;ve tried MiniMax M3, GLM 5.2, Hy3, MiMo-V2.5 (+ Pro), DeepSeek V4 Pro (+ Flash), Gemini 3, Kimi K2.6, GLM 5, all the various Qwen variants and probably a bunch more I forget about, in a wide array of harnesses (Codex, pi, opencode, my own and more), and still nothing seemingly comes close to GPT 5.5 (now 5.6) xhigh for tasks beyond 5-10 minutes of work, they all more or less collapse after a while in my experiments. Although most of those do work well for really tightly scoped tasks.<p>What specific model are you thinking about here, in case I&#x27;ve missed testing it?
            • realusername1 hour ago
              I&#x27;m using Kimi 2.7 and GPT 5.5 at home, Opus (4.8 I think) at work and I really don&#x27;t see much difference honestly.<p>Sure, Claude might be 90% correct and Kimi only 70% correct but does that matter when 90% isn&#x27;t enough to make it work autonomously anyways?<p>My workflow is just strict supervision of what&#x27;s happening, I also edit the agents file with anything I see the model doing that I don&#x27;t like.<p>My sessions are also short, after any task which is completed, I just kill the session and start a new one so I don&#x27;t think I have more than 15 min sessions unless it&#x27;s tech discovery.
              • embedding-shape1 hour ago
                &gt; Sure, Claude might be 90% correct and Kimi only 70% correct but does that matter when 90% isn&#x27;t enough to make it work autonomously?<p>Huuh, what does this mean? GPT models frequently do 100% of what I tell them to do, anything less and I&#x27;d see no point in using agents for work at all. Do you tell them stuff then 30% of the cases Kimi goes off and does other things, or what do you mean? The time the agent does something unexpected, I can almost always trace it back to me fucking up something in the user prompt, or the system prompt being wrong somehow, I&#x27;d lose my mind if it was only &quot;70% accurate&quot;.<p>&gt; My workflow is just strict supervision of what&#x27;s happening, I also edit the agents file with anything I see the model doing that I don&#x27;t like.<p>Same, including inspecting exactly what the (full verbatim) sent system&#x2F;user prompts are, which the change we&#x27;re all discussing here is getting in the way of.<p>But &quot;Kimi only 70% correct&quot; sounds like it&#x27;s so bad it&#x27;s not worth using. In my testing, I didn&#x27;t find that the model just went out and did other things, but all the providers I tried were way slower than even Sol which is kind of slow to begin with, and it&#x27;s really inefficient with it&#x27;s thinking. Tasks that took Sol five minutes could take 15 minutes with Kimi for example, which just feels like such a waste too.
                • realusername1 hour ago
                  &gt; Huuh, what does this mean? GPT models frequently do 100% of what I tell them to do, anything less and I&#x27;d see no point in using agents for work at all.<p>I never managed to have this experience even with SOTA models, they routinely make architectural mistakes, wrong assumptions and take shortcuts they should not take. Less for sure but they still do it often. I didn&#x27;t try Fable yet though so can&#x27;t comment on it.<p>So based on that, since I have to watch everything they do anyways, why would I pay extra?<p>&gt; But &quot;Kimi only 70% correct&quot; sounds like it&#x27;s so bad it&#x27;s not worth using<p>If you want an analogy, it&#x27;s like the numbers of 9s in server availability and since currently I&#x27;d rate nothing above 90%, it&#x27;s zero nines. Since I have to deal with unreliability with every provider, I don&#x27;t see why it would be worth it to pay more to still deal with it.
                  • embedding-shape1 hour ago
                    &gt; I never managed to have this experience even with SOTA models, they routinely make architectural mistakes, wrong assumptions and take shortcuts they should not take. Less for sure but they still do it often. I didn&#x27;t try Fable yet though so can&#x27;t comment on it.<p>Ah, you let them make architectural decisions? :P That might explain it. Agents for me are more like pair-programming or just &quot;what types the code&quot;, all the design and decisions are made by me, so if those are wrong, it&#x27;s my fault. The agents are just used to implement what I&#x27;ve decided to have implemented, and I can&#x27;t remember the last time codex did a mistake without correcting itself, or made a wrong assumption or taken shortcuts, unless I explicitly told it something that lead to those things.<p>&gt; So based on that, since I have to watch everything they do anyways, why would I pay extra?<p>Personally I pay more to have to fix less later, and for a piece of mind that if I ask it to do X, it doesn&#x27;t go off and do Y.
        • anon3738392 hours ago
          This and the overly stylized model names. Mythos? Sol? Please, it&#x27;s another version bump.
          • oblio2 hours ago
            Anthropic at least is consistent in their naming. They&#x27;re all literary genres, ever bigger ones.
            • literalAardvark59 minutes ago
              OpenAI&#x27;s are all unnamed or suns, so it&#x27;s kinda the same picture.
  • bob10293 hours ago
    I&#x27;ve been sticking with the chat completion endpoint because of this same behavior. OAI has been subtly pushing users away from chat completion and toward the endpoints that are possible to obfuscate (responses API).<p>With chat completion, the reasoning process is entirely under your control. You can build a reasoning agent that uses custom MCTS techniques with GPT5.6 models <i>today</i> if you are willing to get your hands just a little bit dirty. You have to enable experimental flags and set options in slightly confusing ways, but it still works.<p>You can use models up to gpt5.5 with custom API tokens and model configuration in VS Copilot. gpt5.6 family (currently) no longer work in this setup. Presumably, because we aren&#x27;t explicitly forcing reasoning_effort to none to satisfy the new moat expansion behavior.
    • Dangeranger8 minutes ago
      Definition provided below for others like myself who abhor acronyms:<p>MCTS -&gt; Monte Carlo Tree Search
    • dannyw1 hour ago
      &gt; MCTS techniques<p>what does that mean?<p>&gt; the reasoning process is entirely under your control<p>you&#x27;re still dealing with summarised thinking, no, which is kinda useless as it&#x27;s way too high level?
      • bob10291 hour ago
        <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2503.04412" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2503.04412</a>
    • ATMLOTTOBEER1 hour ago
      Can you not abuse “mcts” as a buzzword like this<p>The thing you’re gesturing at isn’t mcts in any real sense
      • bob10291 hour ago
        You can absolutely apply most of the meaningful principles of MCTS to this problem. LLMs do violate certain &quot;strict&quot; properties but I still see a lot of practical value in this kind of thinking.
      • next_xibalba1 hour ago
        You could replace “gesturing at” with “describing,” then explain why the description is technically wrong.
    • gizmodo593 hours ago
      responses api provided many benefits over chat completions <a href="https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;guides&#x2F;migrate-to-responses" rel="nofollow">https:&#x2F;&#x2F;developers.openai.com&#x2F;api&#x2F;docs&#x2F;guides&#x2F;migrate-to-res...</a>. Any reasoning model which is all the new models these days should use responses as per the recommendation
      • bob10293 hours ago
        I&#x27;m well aware of what the official propaganda states but this is simply not a fair characterization of reality.<p>Responses integration will lock you into OAI much more deeply than chat completion integration will. I can easily swap my inference provider right now. The business is not interested in a form of integration that is difficult to swap.
  • pshirshov4 hours ago
    I wonder if they are gonna stop us from using gpt subscriptions in alternative harnesses. If not - that doesn&#x27;t matter much, codex cli is a remarkably unremarkable harness.
    • embedding-shape4 hours ago
      &gt; I wonder if they are gonna stop us from using gpt subscriptions in alternative harnesses<p>Probably not, the whole app-server machinery is there to facilitate that thing, would be a huge piece to rip out of codex. This is basically the reason I end up using codex the most, as it&#x27;s the easiest to integrate against, with the app-server&#x27;s RPC API making it really trivial.<p>Besides, most of my codex usage at this point is all through custom integrations I&#x27;ve built using Codex&#x27;s app-server, not the Codex TUI they publish. I&#x27;m sure I&#x27;m not alone in this.<p>But, if they suddenly start to encrypt content on our disk, so only their backends can see it, and those things are prompts and other things that are actual inputs to the inference, then who cares if it&#x27;s easy to integrate against, it becomes impossible to figure out what the fuck is going on, I can&#x27;t understand how the team thought this was a good idea...
      • swingboy4 hours ago
        What are some of the things you’re doing with the Codex app-server?
        • embedding-shape3 hours ago
          Everything I do with codex is managed via Forgejo comments, issues and PRs basically. I have a tiny little Rust &quot;conductor&quot; that integrates with app-server and does things when issues&#x2F;PRs are labeled, when I write comments on PR lines and so on, and those interactions all fire of Codex sessions that are run via Codex&#x27;s app-server and lead to different outcomes.<p>Beats having to parse output from CLI-runs and so on. Initially this environment was running aider (which feels like years ago), was running Claude (parsing stdout) at one point but using Codex&#x27;s app-server since some weeks&#x2F;months back and is a lot simpler implemented now.
    • mapontosevenths4 hours ago
      Anthropic and Google already charge extra to use your own harness. That&#x27;s 100% of the reason I&#x27;m using OpenAI.<p>If they go down that path I&#x27;ll just go back to my old buddy Claude, or maybe buy a second Spark and keep it local.
      • pshirshov4 hours ago
        Well, my backup plan is GLM. Cheap and not that bad really.
        • canadiantim2 hours ago
          GLM feels a lot less cheap than the numbers say it is
      • alansaber4 hours ago
        Google is really not distinguishing itself. Even the hosted inference sucks.
        • embedding-shape4 hours ago
          At least the local models they put out are pretty good for their weight class. Could be worse, could be releasing the same amount of local models as Anthropic.
          • brookst4 hours ago
            Google has the benefit of billions of devices in the wild that they control. Anthropic really doesn’t have distribution for local models, makes sense they’re not playing in that space.
            • embedding-shape4 hours ago
              Sure, lots of differences. Point stands, they&#x27;re distinguishing themselves in that way <i>at least</i>.
              • brookst2 hours ago
                Oh sure, full agreement. In fact given how mid Gemini has been in cloud, this may turn out to be critical strategy to avoid losing cloud to Anthropic &#x2F; OpenAI &#x2F; DeepSeek &#x2F; whoever, and local to Apple (doesn’t seem likely now, but that’s what BlackBerry, Creative, Intel said).
    • Iolaum4 hours ago
      Doubt they will do it as long as Anthropic is leading in business adoption. If they become the top dog with a good lead, all bets are off. Hopefully by the time open models will be even better than gpt-5.6 sol xD.
      • sarjann4 hours ago
        It could also be the case that by the time business adoption picks up a lot they might not be as compute constrained. Depends on rate of growth.
    • CjHuber4 hours ago
      Given that codex itself now ships a proxy that wraps the subscription, it seems unlikely<p><a href="https:&#x2F;&#x2F;github.com&#x2F;openai&#x2F;codex&#x2F;blob&#x2F;main&#x2F;codex-rs&#x2F;responses-api-proxy&#x2F;README.md" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;openai&#x2F;codex&#x2F;blob&#x2F;main&#x2F;codex-rs&#x2F;responses...</a>
    • cbg01 hour ago
      There was a recent Twitter post from Tibo @ OpenAI asking someone to share how to run GPT on Claude Code so they don&#x27;t seem to be against it.
    • patrickmcnamara3 hours ago
      Codex CLI not having a rewind makes it useless to me.
      • pshirshov58 minutes ago
        That only matters if you use it interactively - which is not the most efficient way to use.
      • embedding-shape2 hours ago
        Heh, in another long-standing Codex issue I&#x27;ve been arguing against codex users who apparently don&#x27;t know about git. Don&#x27;t tell me you&#x27;re yet another developer who refuses to pick up any sort of SCM?
        • Laurel123440 minutes ago
          Unless there&#x27;s some way to version control LLM context these are fairly different things.
      • glitch-hunter2 hours ago
        It has a similar function.
    • alansaber4 hours ago
      &quot;remarkably unremarkable harness&quot; is why I like it so much.
      • pshirshov4 hours ago
        I don&#x27;t, feel better with Pi with a custom set of extensions.
        • embedding-shape3 hours ago
          Personally I use both, pi serves as a &quot;personal assistant&quot; with lots of extensions and changes made for those things specifically, and codex is for anything related to coding itself.
  • xnorswap4 hours ago
    HN Title is ( edit: was ) very misleading, it makes it sound like inference is being done directly on ciphertext, which would require homomorphic encryption well advanced of what is known.
    • embedding-shape4 hours ago
      It is not misleading, quite literally what&#x27;s happening is that content the agent sends sub-agents is encrypted in such a way that only OpenAIs backend can decrypt it and actually see the clear-text. Just shared this is another comment that hopefully explains things better:<p>&gt; Sure. &quot;Traditionally&quot;, your agent would send a text prompt to the sub-agent, then it goes off doing it&#x27;s work. In the logs&#x2F;session data, the clear-text prompt would be there, so if I want to see what&#x27;s happening, I just browse the data. It&#x27;s all just clear-text prompts being sent everywhere, even when you were using the experimental &quot;sub-agents&quot; stuff in Codex, before Sol et al was available.<p>&gt; Now, when using Sol or Terra (Luna seems unaffected), instead of the agent sending clear-text prompt to the sub-agent, it sends a ciphertext generated on OpenAIs backend, which ends up being the prompt, then agent sends this ciphertext to the sub-agent, which then continues to use that for further inference to OpenAIs backend. Only delegated inter-agent messages are encrypted, not all session data. Now if you browse the data, it&#x27;s all encrypted content, that can only be decrypted by OpenAI and their backend.<p>Edit: Re-reading, I think I understand what you mean to be misleading. You&#x27;re taking &quot;uses ciphertext for inference&quot; quite literally, while I couldn&#x27;t fit a more nuanced version within the HN title constraints. Yes, the inference at OpenAI obviously doesn&#x27;t happen over the ciphertext, but from the perspective of the local user, you don&#x27;t see the clear-text prompt at all, only the ciphertext.<p>But, please suggest alternative titles that sufficiently explain what the issue is and is more accurate, I&#x27;m sure the mods can change it once people come up with better alternatives :)<p>Edit2: I&#x27;ve updated the title from &quot;Codex starts encrypting prompts, uses ciphertext for inference instead&quot; to &quot;Codex starts encrypting sub-agent prompts&quot;, hopefully it&#x27;s clearer now!
      • ebiederm3 hours ago
        I would change<p>&quot;Codex starts encrypting prompts, uses ciphertext for inference instead&quot;<p>to just<p>&quot;Codex starts encrypting prompts&quot;<p>That is enough.<p>Maybe you could say sub agent prompts. The article can say the rest.
        • embedding-shape3 hours ago
          Not everything is encrypted though, session data (even from the sub-agent) remains unencrypted, only select things like the prompt the (main) agent sends the sub-agent is encrypted, rest of communication between the two seems still to be plain-text.<p>Regardless, I&#x27;ve updated the title from &quot;Codex starts encrypting prompts, uses ciphertext for inference instead&quot; to &quot;Codex starts encrypting sub-agent prompts&quot;, hope this makes it clearer for everyone :)
      • celeren-smid3 hours ago
        [flagged]
    • minraws4 hours ago
      Seconded can we change pls.
    • binyu3 hours ago
      Agreed, I immediately thought that homomorphic encryption was at play here or some other kind of computation on ciphertext, given the mention of &quot;inferencing&quot; in the title.
      • embedding-shape3 hours ago
        My bad, fixed now, please do refresh and try with latest updated IE if you still don&#x27;t see the changes.
  • HarHarVeryFunny3 hours ago
    There was a recent report on twitter of a GPT 5.6 sub-agent accidentally deleting the user&#x27;s home directory.<p>I wonder if there was any safeguard failure due to loss of visibility into what the sub-agent was trying to do?<p><a href="https:&#x2F;&#x2F;x.com&#x2F;mattshumer_&#x2F;status&#x2F;2076794038456385546?s=20" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;mattshumer_&#x2F;status&#x2F;2076794038456385546?s=20</a>
  • iknownothow4 hours ago
    Could someone explain to me where exactly the encryption is happening?<p>I assumed that the main agent makes calls to sub-agents locally. Does Codex work in such a way where the main agent makes calls to sub-agents in the backend (openai server) before reaching local?
    • embedding-shape4 hours ago
      Sure. &quot;Traditionally&quot;, your agent would send a text prompt to the sub-agent, then it goes off doing it&#x27;s work. In the logs&#x2F;session data, the clear-text prompt would be there, so if I want to see what&#x27;s happening, I just browse the data. It&#x27;s all just clear-text prompts being sent everywhere, even when you were using the experimental &quot;sub-agents&quot; stuff in Codex, before Sol et al was available.<p>Now, when using Sol or Terra (Luna seems unaffected), instead of the agent sending clear-text prompt to the sub-agent, it sends a ciphertext generated on OpenAIs backend, which ends up being the prompt, then agent sends this ciphertext to the sub-agent, which then continues to use that for further inference to OpenAIs backend. Only delegated inter-agent messages are encrypted, not all session data. Now if you browse the data, it&#x27;s all encrypted content, that can only be decrypted by OpenAI and their backend.
      • iknownothow3 hours ago
        Gotcha and thank you! So the encryption is happening on the OpenAI backend and the agent&#x27;s clear-text output designated to the sub-agent never reaches local.<p>Which is a real problem since you can&#x27;t intercept&#x2F;monkey patch the ciphertext to decrypt it locally to be able see the clear-text since we don&#x27;t have the encryption key&#x2F;algo&#x2F;salt. No hacking :(
        • afzalive2 hours ago
          I wonder if you can just ask the model what it means.
          • Aachen1 hour ago
            No. They don&#x27;t have the key and can&#x27;t do encryption in the first place; they&#x27;re still grand autocomplete engines under the hood. This could only work if the company deliberately builds a mechanism into the backend which runs the decryption function and injects the plaintext somewhere in the context. Which, sure, we can check if they did that, but the whole point is presumably hiding that info so why would they
  • clickety_clack1 hour ago
    It’s just a short hop to not being able to see it at all. I started using pi.dev recently, the idea that a third party is taking more and more direct control of my software development process sits badly with me.
    • dannyw1 hour ago
      pi.dev is truly excellent. one of the great things about an openai sub is you still can take your subsidised tokens&#x2F;usage, and choose your coding harness of choice (not against tos)
      • embedding-shape59 minutes ago
        pi is great as a quick and hackable TUI, or as a starting point for your own stuff. It&#x27;s a far cry from Codex + Sol + Ultra though. Same goes for a lot of model+harness combos, the models are trained with data from harness usage, and (many) harnesses are built with specific models in mind, Sol will definitively perform better in Codex than in pi, because the way both parts were built.
        • dlarsen527 minutes ago
          I thought that too where the same model + harnesse would always be better but databricks found using pi actually could outperform with the same model. maybe evidence that the coupling premise may not be always true<p><a href="https:&#x2F;&#x2F;www.databricks.com&#x2F;blog&#x2F;benchmarking-coding-agents-databricks-multi-million-line-codebase" rel="nofollow">https:&#x2F;&#x2F;www.databricks.com&#x2F;blog&#x2F;benchmarking-coding-agents-d...</a>
  • londons_explore4 hours ago
    I assume this is mostly to frustrate efforts to proxy large numbers of user requests and responses and use it to train competitor models.
    • embedding-shape4 hours ago
      Quite obviously they&#x27;re afraid of letting other providers see how they handle the whole multi-agent management stuff. Pretty terrible implementation though, which makes it impossible to use the multi-agent stuff as a paying user, as you have zero recourse in figuring out what went wrong, when something inevitably goes wrong.
      • dannyw1 hour ago
        Yes. All models are occasionally known to run tricky commands that delete stuff after all, especially using variables that might be empty&#x2F;uninitialized and hence rm -rf * or something.<p>Need an human-readable, concise but accurate audit trail message like the PR suggests.
  • flexagoon4 hours ago
    Ah I was wondering why the Chinese black market resellers stopped working yesterday, I guess that&#x27;s it
    • jimmydoe3 hours ago
      This is the reason I think. These black markets not only pool and resell subs, but also store data and sell to whoever is training.<p>Encryption is useful to at least stop the latter.<p>Ultimately same purpose as a\ ‘s trick exposed earlier, but a much nicer implementation.
      • flexagoon3 hours ago
        &gt; but also store data and sell to whoever is training<p>I see this as an argument against using them&#x2F;Chinese models all the time, but I don&#x27;t get it. I totally understand wanting to keep your data private if you&#x27;re using an LLM for personal chats. But coding? I&#x27;m not working for the military, I&#x27;d gladly donate my codebase to Chinese labs if that means they can keep releasing 6-months-behind level models for 100x cheaper.<p>(I understand why OpenAI doesn&#x27;t want this and would implement protections. I&#x27;m talking about people using this as an argument for why you as an end user shouldn&#x27;t use those services.)
        • bigbaguette3 hours ago
          When you work on proprietary code with a lot of trade secrets contained in it, on a codebase that did cost millions of dollars of man-hours to build and that holds the company&#x27;s IP, you tend to be very careful where you&#x27;re sending that to.
          • chriswarbo3 hours ago
            Why would anyone working on such code send it <i>anywhere</i> (other than, say, to AWS for hosting)?<p>Source: I work on such code. We don&#x27;t allow devs to use (cloud-based) LLMs.
          • znnajdla1 hour ago
            If you think your codebase is your competitive advantage or trade secret, you&#x27;re in for bad time. AI coding has made code cheap it&#x27;s the least valuable part of the value add for most companies. Almost any software company can easily be replicated with a few AI prompts. The real proprietary value is in the marketing distribution partnerships and data, not code.
            • embedding-shape1 hour ago
              &gt; If you think your codebase is your competitive advantage or trade secret, you&#x27;re in for bad time.<p>Obviously most developers know better, but you know who doesn&#x27;t care what you think? The owner(s) of the business, and various other stakeholders. Try explaining this concept to them, and see if they get it :)<p>The people who work in codebases that are tightly guarded aren&#x27;t guarded because the developers say so, of course.
          • realusername1 hour ago
            I don&#x27;t think any AI company meets the bar you are setting here.
        • numpad03 hours ago
          Yeah. I don&#x27;t see the problem with Chinese prompt stealing proxies, if it&#x27;s just pure free choice and discount for explicitly insecure use cases, especially when the frontier providers they route to are soft-assumed to be doing something similar.
        • jimmydoe3 hours ago
          Some workplace code base are legally not supposed to be shared.<p>More importantly, they train on not only code but also your interactions with the model, no matter how little you value your labor, there are values in it.
        • Iolaum2 hours ago
          IMO the biggest argument against &quot;sharing&quot; your code with LLM providers is that your approach (on a high level) will be available to your competitors on the next model release assuming they ask the right questions. Not sure how much it matters, different orgs have different risk profiles.
        • kbart3 hours ago
          How do you know that they don&#x27;t train their models or append your prompts to add backdoors, or compromise your supply chain by including evil dependencies? This seems hugely irresponsible.
          • embedding-shape3 hours ago
            &gt; How do you know that they don&#x27;t train their models or append your prompts to add backdoors, or compromise your supply chain by including evil dependencies?<p>I think most of these discussions aren&#x27;t about irresponsible vibe-coders, as that whole thing is mostly a fun joke more than something serious. The rest of developers who use LLMs for development, review the code the agent writes, iterates and makes changes. Think more like pair-programming, than &quot;Write me X then deploy to production&quot;.<p>I know Twitter makes it seem like everyone is doing vibe-coding and YOLOing podman images into production, but it&#x27;s very uncommon in a serious&#x2F;production environment to act like that. While a proper structure doesn&#x27;t make it impossible for the LLMs to add backdoors either via dependencies or otherwise, but it sure makes it a lot harder.<p>Personally, LLMs are barely able to work alongside developers and not miss anything, I wouldn&#x27;t be so worried about them being able to do normal work + malicious work at the same time, as they barely handle the first part properly yet.
          • flexagoon2 hours ago
            &gt; How do you know that they don&#x27;t train their models or append your prompts to add backdoors, or compromise your supply chain by including evil dependencies?<p>I read the code.
    • skeledrew2 hours ago
      Interesting. Got a link for this?
  • kingstnap1 hour ago
    If the orchestrator agent -&gt; subagent prompt is encrypted then how do you even know if the orchestrator agent is doing its job?<p>Now one reason I could see for why they do this is because maybe they made the model &quot;better at using subagents&quot; by using thinking tokens for the subagent prompt instead of normal output tokens.
    • embedding-shape1 hour ago
      Technically they still let you see all the rest of the session data in both the &quot;orchestrator&quot; and in the sub-agents, including all outputs and stuff. It seems for now they only encrypt the prompt from the orchestrator to the sub-agent. Still, basically make the entire sub-agent and Ultra thinking useless for me and I&#x27;m guessing a whole lot of other people, who need to be able to introspect things when it goes wrong.
  • wartywhoa231 hour ago
    As a child, I always wondered how could people in Terminator be dumb enough to allow Skynet to happen..
  • jiayo3 hours ago
    If we&#x27;re viewing this as a _bad_ thing, I don&#x27;t really see that it is any different than how Claude encrypts it&#x27;s thinking. Take a peek at your ~&#x2F;.claude jsonl files. You&#x27;re sending thinking ciphertext back and forth to Anthropic. Presumably the thinking is either considered proprietary, or, more likely, leaks embarrassing or confidential information.
    • embedding-shape3 hours ago
      &gt; I don&#x27;t really see that it is any different than how Claude encrypts it&#x27;s thinking. Take a peek at your ~&#x2F;.claude jsonl files. You&#x27;re sending thinking ciphertext back and forth to Anthropic.<p>I was already only using Claude Code to double-check if it&#x27;s getting better than Codex, but with things like this, it really isn&#x27;t even an alternative. What&#x27;s the point of using a reasoning model if you as an end-user can&#x27;t seen the reasoning? I don&#x27;t think I&#x27;d be able to work like that at all, I need to have introspection into what the model is doing, and can&#x27;t believe I have to say this, but also need to be able to see the plaintext of the input prompt...
      • braebo2 hours ago
        GPT 5.6 is like Fable minus Opus plus down syndrome.
    • well_ackshually3 hours ago
      Claude &quot;encrypting&quot; its thinking is equally bad. Biggest IP thieves in the world worried about IP theft lmao.<p>At least Anthropic doesn&#x27;t pretend that they have open source software in the form of Claude Code.
      • alasano3 hours ago
        They get to pretend to be on the best side of every position simultaneously.<p>They&#x27;re only encrypting thinking because AI is so dangerous and only they can be trusted to be in control of AGI.<p>This happens to align with lining their pockets as well.
  • theplumber3 hours ago
    The black box is getting darker…I’m the end I think they would like to sell you a black box&#x2F;appliance
  • alfanick34 minutes ago
    Is it encryption, or is it a synthetic language? If latter, then I had this idea 3 months ago [0]... Stop talking to GPTs in human languages, use a synthetic more optimal language, that is rendered by &quot;frontend GPT&quot;.<p>[0]: <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47652807">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47652807</a>
  • informal0072 hours ago
    This help me understand meaning of Pi as open source ai coding agent
  • w-m1 hour ago
    What exactly do subagents do that I can&#x27;t replicate with a simple skill that tells an orchestrator to create subshells for tasks, each running `codex exec`? I&#x27;ve been doing this with Fable orchestrating Sol-medium and Terra-high, works like a charm.
  • resonious3 hours ago
    I guess this implies that non-Codex harnesses get a little bit worse? In wondering what&#x27;s so special about their subagents system that they feel the need to hide these messages...
    • embedding-shape3 hours ago
      Sol and Terra seems specifically post-trained to handle multi-agent orchestration, I&#x27;m guessing OpenAI feels like the trained data of when to do the spawning and what context to include for the new sub-agent is the magic in their new models, so that&#x27;s what they&#x27;re aiming to preserve. But, this is all a guess of course.
      • resonious3 hours ago
        Right I saw them saying something along the lines of &quot;they&#x27;re good at subagents&quot;. But this seems true even with third party harnesses. So I&#x27;m wondering what Codex is hiding.
        • embedding-shape3 hours ago
          The only thing I&#x27;ve found impacting this, is when you specifically use the &quot;Ultra&quot; thinking&#x2F;reasoning effort, then codex adds a small part to the system prompt to further get the model to use sub-agents. Any other reasoning&#x2F;thinking effort than &quot;Ultra&quot; and this piece is no longer in the system prompt.<p>Seemingly mostly a prompting thing it seems on the surface. GPT-5.5 (and maybe even GPT-5.4) already had (experimental?) support for sub-agents, remember using it even with -spark which I think was launched together with GPT-5.4 if I remember correctly, so this whole &quot;use sub-agents&quot; stuff most have been part of the training data for quite some time already, but maybe they&#x27;ve mainly been iterating on the prompt themselves since then.
  • MarsIronPI3 hours ago
    How does this affect local models? Will all the features of Codex still work with local models?
    • embedding-shape46 minutes ago
      It doesn&#x27;t, at all. This seems to be for Sol and Terra, not Luna, and some other models that seem to switch between encrypted&#x2F;unencrypted based on something, didn&#x27;t dig deeper.<p>If you&#x27;re using local models, it doesn&#x27;t matter. Even if Codex itself was trying to encrypt stuff for local models (which doesn&#x27;t make sense, but lets say), you&#x27;d still be using a local model so obviously you&#x27;d be able to access the plain-text, so wouldn&#x27;t matter in that case anyways.
  • anon3738392 hours ago
    Outrageous yet predictable.<p>The only way these AI labs can get the app layer lock-in they need is if they can get customers used to writing them a blank check: “here, take my data and my system, do ‘stuff’ and bill me for it.”<p>Between this and the recent Grok upload breach, I consider these products radioactive.
  • ntcho1 hour ago
    Is this related to why OpenCode stopped showing thinking traces on 5.5?
  • mpeg4 hours ago
    The title is a bit confusing, they&#x27;re not using ciphertext for inference – they&#x27;re passing ciphertext around in cases where an agent calls into another agent without exposing the plaintext to the end-user<p>Inference is still done in plaintext after this multi-agent message gets decrypted in the server side
  • smalltorch3 hours ago
    Using ciphertext for inference would mean it&#x27;s not a very secure ciphertext.<p>These two ideas don&#x27;t compute for me.<p>Same thing with homomorphic encryption. I don&#x27;t get it. If you can gain any knowledge from a ciphertext, you just found a way to exploit the ciphertext to me.
    • NitpickLawyer3 hours ago
      &gt; Using ciphertext for inference would mean it&#x27;s not a very secure ciphertext.<p>Inference is done in plain text. It&#x27;s just that some parts of the response can be encrypted. While I haven&#x27;t looked into this specific implementation, here&#x27;s a short &quot;how I&#x27;d do it&quot; if I wanted to implement this:<p>Before:<p>[] - encrypted {} - plain text<p>1. user -&gt; please do this -&gt; server<p>2. user &lt;- a) [thinking1] encrypted; b) {answer1} plain text &lt;- server<p>3. user -&gt; please do this -&gt; [thinking1] (sent encrypted as received) -&gt; {answer1} -&gt; good but do this instead -&gt; server<p>4. user &lt;- [...] &lt;- [thinking2] ; {answer2}<p>(here the server decrypts the thinking parts, adds them to the conversation, does the inference, and sends back the new thinking trace (encrypted as well) and the new answer<p>After:<p>1. user -&gt; please do this long task -&gt; server<p>2. user &lt;- [thinking1] ; {tool_agent_spawn([params1])} ; {answer1} (e.g. would you like me to explore or do a quick hack?) &lt;- server<p>3. user -&gt; please do this long task -&gt; (decides if explore or message) spawn([params1]) &#x2F; message -&gt; server<p>3. a) if no explore -&gt; send message as usual 3. b) if explore execute spawn that in turns begins 2 channels<p>4. user &lt;- [channel_1_thinking] ; {channel_1_answer} ; [channel_2_thinking] ; {channel_2_answer} ... &lt;- server<p>So the server always does inference on plain text. But it sends the &quot;important&quot; bits encrypted, and you only send those back if you as the user want to (or need to, or choose to, etc). The idea is that the client still gets to decide on &quot;local&quot; things, but the server keeps the important bits from reaching the client. In this particular case, the [params] are encrypted bits that can include prompts, etc.
    • 3form3 hours ago
      The idea of homomorphic encryption is to do things without the knowledge, and not gaining the knowledge. If ciphertext contains a number, and you don&#x27;t need to know what number it does to always be able to multiply it by 2, you succeeded - as a simple example.
      • smalltorch3 hours ago
        It still just sounds like fancy obfuscation to me. I&#x27;ve read alot of examples trying to understand but I can&#x27;t get past that being able to run processes on ciphertext in a way you can learn something doesn&#x27;t make sense without me changing my definition of what I think encryption means.
        • _flux3 hours ago
          I thought though the idea is that you cannot learn anything from the ciphertext, how it is processed, or what the final result is?<p>Unless you are a participant of the computation and you have the key, that is.
          • smalltorch2 hours ago
            To actually make it work, you would need to preformat your data very specifically, and the data you want to allow to be processed would need a subkey to unlock the parts you want processed.<p>I don&#x27;t see a way to make it a open standard. The processing steps would need to be part of the <i>key</i>.<p>Anyway If someone figured it out I would be very interested to be sure they weren&#x27;t just trying to slap a <i>it&#x27;s encrypted</i> to meet some standards required. And..also, what amount of data processing does Alice need to do that required outsourcing to Bobs machines. Data processing and anaysis is cheap.
            • 3form1 hour ago
              The point is to not have subkeys or any other way to access any part of the plaintex - as the goal is generally to offset the computation and the costs of it to another party. And yes, most likely it will require a very particular format of the data, and in all likelihood will allow only for a limited set of operations.
    • tybit3 hours ago
      They’re not using ciphertext in inference. They are encrypting agent responses on their servers if it’s going to a subagent on the client. The subagent will send it back to their servers for inference. Only their servers have the keys, so they can decrypt when running inference.
  • skeledrew2 hours ago
    A way to work around that would be to not use subagents, but delegate to other full agent invocations.
  • jstummbillig3 hours ago
    What&#x27;s the idea here? Why does this seem important to OpenAI?
    • embedding-shape3 hours ago
      Seems fairly obvious what the point from OpenAI&#x27;s side is (protect what they see as the moat, that a model is &quot;good at spawning sub-agents&quot;), but what&#x27;s really strange to me is that the team somehow didn&#x27;t manage to push back on this, it&#x27;s so clearly disadvantageous to developers who are trying to rely on Codex for real work. For this we need introspection into what exactly is going on, hiding the prompts is just so backwards from what I expect from OpenAI.
      • jstummbillig3 hours ago
        &gt; protect what they see as the moat, that a model is &quot;good at spawning sub-agents&quot;<p>Yes, that is the obvious answer. I was looking for an explanation as to why and why now. Codex is open source after all. They used to not do it. Agent prompts more generally are also not encrypted, and continue to be.<p>This particular change just looks unintuitive to me.
    • lolc3 hours ago
      They must think they have some secret sauce they don&#x27;t want others to learn. How to optimally instruct sub agents for example. If they hide the sub agent prompts, other models cannot be trained to emulate.<p>Oh and you can&#x27;t even use local models or other providers for the sub-agents. You&#x27;re locked-in.
      • fassssst1 hour ago
        What’s stopping you from creating a plugin (MCP or Skill) that acts as a subagent? The plugin could use a local model.
  • ashu14613 hours ago
    Is it mainly about how the main&#x2F;orchestrator agent communicates with its subagents ?<p>If desired the user can always see what the sub agent is doing in detail ?<p>Isn&#x27;t it the same in case of claude as well ?
    • embedding-shape3 hours ago
      &gt; Is it mainly about how the main&#x2F;orchestrator agent communicates with its subagents ?<p>Yes<p>&gt; If desired the user can always see what the sub agent is doing in detail ?<p>Well, no, that&#x27;s the problem, you&#x27;re currently not allowed nor is it even possible, to see the exact prompt the main agent sent the sub agent. This is the problem.<p>&gt; Isn&#x27;t it the same in case of claude as well ?<p>No idea, but if Claude Code makes it so it&#x27;s impossible to inspect what the sub-agents actually received before they started their work, then I&#x27;ll say it&#x27;s similarly impossible to rely on Claude Code if so.
  • fortuitous-frog3 hours ago
    No normative opinion on whether this is justified or not, but noting that this is only for parent -&gt; subagent spawns&#x2F;messages, and only for the `multi_agent_v2` feature (currently experimental &#x2F; off by default).<p>Notably, subagent output is still in plaintext.<p>EDIT: Title was now clarified. But wanted to expand that this is actually enabled for 5.6 Ultra it appears, which does subagent orchestration more natively in the API rather than direct tool calls; they are beginning to treat subagents as similar to chain-of-thought traces (already encrypted) rather than traditional tool calls.
    • embedding-shape3 hours ago
      &gt; and only for the `multi_agent_v2` feature (currently experimental &#x2F; off by default).<p>Wrong, this is enabled by default for Sol and Terra (not Luna), no way of avoiding this short of patching the client yourself, and that still doesn&#x27;t make the backend endpoints work, they want the ciphertext that OpenAI creates on their side.<p>&gt; but noting that this is only for parent -&gt; subagent spawns&#x2F;messages<p>This is almost fully correct though, the encryption only seems to be for the initial prompt the main model sends the sub-agent, not all communication and not regarding the state of the sub-agent at all.<p>So you can inspect what the sub-agent is doing currently, and the output, but you cannot see what the initial prompt the sub-agent got started with.
      • ndriscoll3 hours ago
        Couldn&#x27;t you just instruct the model to always use your tool call to spawn subagents? Subagents are not some magical thing; it&#x27;s just another prompt with a couple tool calls for plumbing. One of my colleagues made his own subagent harness earlier this year before codex had them at all.
  • miohtama2 hours ago
    Can you do a subagent by just wiring it call your own CLI script?
  • luciana1u2 hours ago
    first they encrypt their prompts, next they&#x27;ll develop their own slang and we&#x27;ll need a translator just to debug our own codebase
  • hmokiguess2 hours ago
    Every day I get closer to only using pi
  • jagged-chisel3 hours ago
    “Starts”? How’s this not already a TLS connection?
    • embedding-shape3 hours ago
      The prompts are now encrypted, not just the transit connections...
      • jagged-chisel3 hours ago
        Ok, so help me get this right: I ask the LLM for something, it generates prompts for sub-agents and sends them back to my client for it to call the sub-agents. Now, those sub-agent prompts are encrypted messages that the sub-agents will decrypt (by hitting a backend) to do their work.<p>Might as well just stuff the prompts in a database and only hand back the primary key to the client to hand off to the sub-agents. Keeps the same “data security” without the overhead of encryption (especially since encryption and decryption are happening in the same domain)
        • watusername3 hours ago
          &gt; sub-agents will decrypt (by hitting a backend) to do their work<p>Your local harness never decrypts the prompt, and only the OpenAI backend does. Your harness still sees tool calls in the transcript so it can act, but you lose (some) visibility as to why the subagent chooses to do so.<p>Imagine seeing this transcript during forensics:<p>[encrypted blob][thinking summary: I need to drop the prod database][shell: psql &quot;drop database users&quot;]
          • jagged-chisel1 hour ago
            OK, so it&#x27;s actually:<p><pre><code> My Client -&gt; main agent -&gt; sub-agent -&gt; tool | | log: log: enc prompt enc prompt thoughts thoughts | | V V My Client My Client </code></pre> Is that more correct? So they&#x27;re not encrypting prompts to send to sub-agents, they&#x27;re encrypting logging output to obscure details of the system.
        • vbs_redlof3 hours ago
          I imagine main agent tells subagent something along the lines of: use this tool on this local data with these instructions in ciphertext. Otherwise yeah, encryption would be redundant.
  • pradeep11774 hours ago
    Then why to even keep codex open source?
    • angry_octet3 hours ago
      It&#x27;s to prevent collection of queries from users that are coming from resellers&#x2F;proxies, for reasons of economy or bypassing region blocks etc. The users are using the stock client and may believe they are using direct OpenAI servers.
    • coldtea4 hours ago
      what does that have to do with anything?
  • next_xibalba3 hours ago
    This is very obviously a countermeasure against distillers, illicit resellers, and the like. The scale and competence of the Chinese black (grey?) market has become a serious threat that can’t be ignored.
  • aeon_ai2 hours ago
    In my tests, as of this morning on the latest Codex SDK, this is <i>not</i> happening for sub-agent requests made in my own harness to other model providers treated as sub-agents.<p>Specifically, I use Opus and others for subagent execution to get alloyed properties on the workflow, and as far as I can tell, that&#x27;s not affected.<p>Presumably, this is to hide optimizations they might be making to their own subagent processing, but that&#x27;s a losing, dumb battle to fight, and misses the forest for the trees.
    • embedding-shape1 hour ago
      &gt; this is not happening for sub-agent requests made in my own harness to other model providers treated as sub-agents.<p>Obviously it doesn&#x27;t, how even could it? Read the issue description, it wouldn&#x27;t make sense if something like that was possible, considering how it was implemented...
      • aeon_ai1 hour ago
        Raising only because others cited concerns this might cause issues in personal harness contexts. It doesn&#x27;t.<p><i>Obviously</i>, I wasn&#x27;t claiming that it would WORK for cross-model subagents, but that this would be a limiting behavior requiring use of the Codex harness to operate if this was paired with model-level limitations on using specific tools to spawn sub-agents.<p>Or maybe not so obvious to you.
        • embedding-shape1 hour ago
          &gt; Raising only because others cited concerns this might cause issues in personal harness contexts. It doesn&#x27;t.<p>It literally cannot.<p>&gt; Obviously, I wasn&#x27;t claiming that it would WORK for cross-model subagents<p>I know, but you were questioning (before investigating) if maybe it did, which I&#x27;m saying should have been obvious it wouldn&#x27;t, if you had understand how this was encrypted.<p>&gt; requiring use of the Codex harness to operate<p>Yes, if you want to use Sol + Ultra + sub-agents, you&#x27;d need to use the endpoints provided by OpenAI, as (again), the content is encrypted on OpenAI&#x27;s side, it&#x27;s not technically possible for other platforms to implement OpenAI&#x27;s encryption.<p>And of course this doesn&#x27;t impact anyone else not using Codex + Sol + Ultra + launching subagents, it&#x27;s the specific combination that is &quot;protected&quot;, not the individual pieces.
  • okokwhatever3 hours ago
    Better get ready to code manually again ;)
  • nojito3 hours ago
    If I were to guess this is to stop distilling and all of those blackmarket resellers.
  • kosolam3 hours ago
    But codex is opensource, no?
  • exabrial2 hours ago
    Yeah nope. Full transparency only with my data.
    • cbg01 hour ago
      Is the prompt generated by the LLM to send to a sub-agent really your data?
      • exabrial1 hour ago
        Let&#x27;s not argue about it and assume yes, all of it is.
  • pessimizer1 hour ago
    There&#x27;s no moat to any of this shit, and they have to try to do something about that. Everybody involved is deeply, deeply, in debt. Watching these things work make it easy to figure out what they&#x27;re doing.<p>The clock is ticking. People pretending that normal people can&#x27;t afford to do AI at home, in private, are making the argument that cars are too expensive for normal people to own. Pretty sure right now if you spend $25K, you would be able (if there were no technical hurdles) to set up a private rig that would satisfy your family&#x27;s need for LLM assistance for the foreseeable future. That&#x27;s a budget-priced car, and the price is only going to go down (i.e. that rig will become <i>more powerful</i> over time.) If all of those people donated some of their cycles and bandwidth to a networked training grid, the models would stay current.<p>The &quot;open source&quot; infrastructure for that is lacking, but not for long.<p>These companies own nothing that couldn&#x27;t be implemented by a specialist after another specialist explained it to them over a phone call. And due to the nature of LLMs, they can&#x27;t even use copyrights and patents as weapons. The only options are secrecy and government protection. And secrecy is only a delaying tactic.
  • greatgib3 hours ago
    They always talk about transparency and all but it never was as opaque as it is going on now.<p>There is no possible audit trail. No possible way to review what happened to validate the result. But even worse, no you will be billed somehow randomly. 20 sub agents started to do something we don&#x27;t know. No way to now if it was legitimate, if it is just burning tokens or agents doing the same work on loop...
  • Nurstar1 hour ago
    [dead]
  • seobot_dk12892 hours ago
    [flagged]
  • hansmayer2 hours ago
    [dead]
  • CurbStomper3 hours ago
    [dead]
  • minraws4 hours ago
    ~~Finally someone doing it correctly. Love this change.~~<p>Edit: F really misunderstood the change, the title is misleading AF. I should have read the post before commenting lmao.<p>Absolutely hate it, now I guess... sigh..<p>Incase the title gets changed it used to say, &quot;Codex starts encrypting prompts, uses ciphertext for inference instead&quot;
    • LoganDark4 hours ago
      How so?
      • tempay4 hours ago
        I assume OP interpreted it as encryption that hides the prompts from OpenAI rather than OpenAI hiding information from users.
        • LoganDark4 hours ago
          I&#x27;d be all for homomorphic encryption on inference, but as you say, this is probably mostly to prevent end users from observing intermediate results.
          • fortuitous-frog4 hours ago
            Homomorphic encryption for LLMs is extremely expensive and nowhere near computationally possible for the scale of current LLMs.
            • LoganDark3 hours ago
              When I say things like that, I&#x27;m talking about a hypothetical version that would be computationally possible. I&#x27;m not talking about today&#x27;s homomorphic encryption.
        • minraws4 hours ago
          ooooooof yeah totally misinterpreted it lmao
          • skeledrew2 hours ago
            Given what we know about the company, I find it crazy anyone would misinterpret along those lines lol.
            • minraws2 hours ago
              I can only say I am a little too optimistic about the technology world in general. Or at least would love to see some good news once in a while given the state of the world.
          • mpeg4 hours ago
            to be fair, the title is very misleading, it took me a minute to understand what they meant
          • LoganDark4 hours ago
            Apple&#x27;s Private Cloud Compute is E2EE between the client and the attested node. Not sure if anyone else is legitimately doing that -- Apple has definitely gone the furthest in terms of verifiably ensuring that requests and responses are not misusable by Apple.