2 comments

  • echoangle11 minutes ago
    At that point you can just paste a screenshot of your doc into word and celebrate.<p>Also, the mitigation can probably be fooled with ligatures since they are only verifying the letters alone as far as I skimmed.<p>I don’t even understand the threat model. Is my opponent in a court case going to use this on the PDF they give the court? Surely the judge will be pretty annoyed since you can’t even ctrl+f in the files then.
    • piker10 minutes ago
      That&#x27;s true for the full obfuscation, but not for the replacement. For replacement there&#x27;s really nothing like it. We just shared the full obfuscation as just a PoC.<p>[Edit: The point here is not to prove some massive &quot;gotcha&quot;, but rather demonstrate that there are a whole class of vulnerabilities that these pipelines are subject to. There will be follow-up posts that pack much more punch.]
      • echoangle8 minutes ago
        Assuming you’re the author since you also posted it: I just stealth-edited my comment, could you maybe talk about the threat model a bit more? I am not a lawyer so I don’t really see when I would want to do this.<p>Also, I hope the „lame exploit“ I just edited out was not too offensive, it’s always great when people try to find attacks to make systems more safe.
        • piker2 minutes ago
          Absolutely, and we definitely agree this particular attack is &quot;lame&quot; in the sense of not allowing CVE, etc.<p>But, we&#x27;re working on a lot of these (as we encounter them in developing Tritium), and the point really is just to demonstrate that LLMs can be blind to ineffective implementations of the specs and other tricks.<p>As mentioned in the accompanying LegalQuants post, we see a lot of these available in the pipelines of applications like Claude for Legal, Harvey, Legora and others.<p>The most nefarious case here requires crafting a number of custom fonts to do character-swapping. It&#x27;s less discoverable but may be sanctionable to your point.<p>But bear in mind this particular &quot;attack&quot; was vibe coded in a day or two and most of the frontier models fail to pick up on it. As &quot;AI native&quot; firms come on line, and aim to be increasingly end-to-end automated, these will become real legal issues.<p>And there will be a lot of them available.
  • mproud11 minutes ago
    Someone could also just make a font file that swaps all of the characters around. So like an A looks like a Z, and a Z looks like an A.
    • piker10 minutes ago
      Covered in the post! It&#x27;s the more aggressive approach for sure.