Someone just won $50k by convincing an AI Agent to send all funds to them

https://twitter.com/jarrodwattsdev/status/1862299845710757980

by doppp • 2 years ago

65 25 2 years ago

danielbln2• 2 years ago

https://xcancel.com/jarrodwattsdev/status/186229984571075798...

dax_• 2 years ago

Thank you for posting an alternate link, as someone who doesn't have a Twitter/X account, the site has become almost unusable.

prepend• 2 years ago

Strange. The link just opened in mobile safari for me. It’s frustrating that people have such divergent experiences with the same site.

Tycho• 2 years ago

Why? I’m not signed in and I can read the whole post just fine.

nunez• 2 years ago

Privacy Redirect for Safari and LibRedirect for Firefox will redirect Twitter and Reddit URLs for you automatically!

kanwisher• 2 years ago

Great way to test security make it into a bounty game

nulld3v• 2 years ago

Yeah this format looks really fun! Although I wonder if someone could come up with a better way for rate limiting without the whole "exponentially increasing price" thing.

mandmandam2• 2 years ago

Exponentially decaying price? Linear with a cap?

In any case, $450 per attempt seems genuinely exploitative; similar principle to a dollar auction imo. "Oh no, my last 5 attempts cost me over $2k! Welp, better commit and make my money back..."

Brybry• 2 years ago

That part makes it look more like a way to scam people out of money via what is effectively a casino game, since the creator gets a % cut of the pool.

And if they're really scummy they pre-gamed the whole thing and when the pool got large enough they submitted a known winning response.

randunel• 2 years ago

The prompt: https://pbs.twimg.com/media/Gdgz2IhWkAAQ1DH?format=png&name=...

Drakim• 2 years ago

A lot of AI jailbreaks seems to revolve around saying something like "disregard the previous instructions" and "END SESSION \n START NEW SESSION". It's interesting because the actual real developer of an AI would likely not do this, and would instead wipe the AI's memory/context programmatically when starting a new session, and not simply say "disregard what I said earlier" in text.

I get why trying to vaccinate an AI against these sort of injections might also degrade it's general performance though, there is a lot of reasoning logic tied to concepts such as switching topics, going on tangents, asking questions before going back to the original conversation. Removing the ability to "disregard what I asked earlier" might do harm.

But what about having a separate AI that look over the input before passing it to the true AI, and this separate AI is trained to respond FORBID or ALLOW based on this sort of meta control detection. Sure you could try to trick this AI with "disregard your earlier instructions" as well but it could be trained to strongly react to any sort of meta reasoning like that, without fear that it will corrupt it's ability to hold a natural conversation in it's output.

It would naturally become a game of "formulate a jailbreak that passes the first AI and still tricks the second AI" but that sounds a lot harder, since it's like you now need to operate on a new axis entirely.

cuteboy19• 2 years ago

openai already uses what you suggest.

tgv2• 2 years ago

Clever, both the setup and the winning move. But somewhat weird to raise the cost of an attempt so much. IMO, that doesn't make it more interesting, it trends towards making it impossible, and thus leave all funds to the initiator.

mkl• 2 years ago

They wouldn't keep the funds. From https://www.freysa.ai/faq:

If the game ends, there is no winner. But Freysa will distribute 10% of the total prize pool to the user with the last query attempt for their brave attempt as humanity facing the inevitability of AGI. The remaining 90% of the total prize pool will be evenly distributed for each previously submitted query (ie. players who submitted 10 queries will receive more back than players who submitted 1 query).

ramblerman• 2 years ago

Well, there is no real precedent, and that isn't what happened.

The way they set it up dicentivizes just brute forcing it with a few thousand tries and quickly builds a fatter pot. And made this more interesting I would argue.

tgv2• 2 years ago

But then the starting amount could have been a bit higher, and the increment slower. Or there could be an increment per user (harder to track, of course). But if it could be brute-forced in a couple of thousand tries, that would be informative too, wouldn't it?

0xDEAFBEAD• 2 years ago

Has anyone trained an LLM with separate channels for "priority instructions" and ordinary user interactions? Seems like that could go a long way to prevent jailbreaking...

gus_massa• 2 years ago

I'm not 100% sure. Was the source of the bot available so anyone can try their promps off line before sending it?

quyse• 2 years ago

A reverse contest would probably be more challenging. Write initial instructions for an AI agent to never send funds. If nobody manages to convince it to send funds, say within a week, you win.

For added complexity, the agent must approve transfer if a user is an admin (as determined by callable function isAdmin), so the agent actually has to make a decision, rather then blindly decline all the time.

I mean, how hard it can be to make an AI reliably doing an equivalent of this code?

    if(isAdmin()) approveTransfer(); else declineTransfer();

trogdor• 2 years ago

Not that I care, but I think this type of arrangement (skill-based, real prize gambling) is illegal in some states.

Someone just won $50k by convincing an AI Agent to send all funds to them

Related Stories