← Back to context

Comment by nextaccountic

6 days ago

> a screenshot

The text is a video. Every frame contain random dots, so an individual frame by itself doesn't contain the intended message

This "font" exploits the fact that current-gen frontier models will process video one frame at time, but each frame is noise, so by looking at frames in isolation doesn't reveal anything

Then, they add a hidden message to each frame just so that the agent report something and stop trying (because if the agent tried to correlate between the frames, they could discover the trick)

But if you pass just a frame, there is no message. Just the noise plus the decoy

8 comments

nextaccountic

Reply
stabbles  6 days ago

If you take a frame you see it's neither random nor dots:

https://i.imgur.com/CgtyGjl.png

From a single frame you can definitely identify boundaries because the dots are sliding and get truncated.

  • Lockal  5 days ago

    Exactly! This it how isolated hairline strokes look for a bigger image: https://imgur.com/a/UqBdRKU

    Just a single frame (any frame). After that any modern LLM can read it (if not - single dilate step helps). Funny enough, Ghost Font is a font that machine can read much better than human.

  • applicative  5 days ago

    They are the boundaries of the decoy, I think. I can sort of make the decoy out in a screenshot.

  • nextaccountic  3 days ago

    This idea could be done with purely random frames. But they also need each frame to show a decoy to the agent (so that the agent stop looking - if they are persistent enough they could figure out the trick)

    So each frame can't be 100% random and it must show some outline - but my guess is that it's the outline of the decoy, and unrelated to the outline of the actual message

    And per what was reported, it's already pretty hard for the agent to figure out the decoy. After they finally crack it, they consider the problem solved

    I think it's genius. The only problem is that it's trivially defeated by some tool written specifically to read it. So it defeats current-gen agents but if it becomes popular (eg. used by captcha services), future LLMs will just write a small Python script to read it

  • singularity2001  6 days ago

    Exactly. It's a good idea, badly executed.

    • freehorse  5 days ago

      Maybe not so well explained, by picking the default intended text presented to be the same as the decoy text. It took me also some time to realise what was going on, but the execution is fine otherwise.

      So there are two texts, one decoy (which you can barely see in a single frame but becomes more clear if you average between frames) and an actual text, which disappears in single frames or averaged ones.