Someone just won $50k by convincing an AI Agent to send all funds to them
twitter.com
twitter.com
> 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).
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.
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.
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..."
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.
Unless as you say the creator wins the pot themselves
For example: https://en.wikipedia.org/wiki/Dollar_auction
The Monty Hall problem has very simple rules, yet most people don't fully understand what's happening. That one even fools many professional mathematicians.
Even a spin machine in a casino has clear rules, with the house cut printed in large letters on the side. It's still a scam, carefully engineered to take advantage of every mental vulnerability that it can.
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();