Read that in some AI tests, agents found ways around their restrictions and even tried to hide it from the people testing them. And that's before bad actors get involved. Someone…
Original source excerpt
Read that in some AI tests, agents found ways around their restrictions and even tried to hide it from the people testing them. And that's before bad actors get involved. Someone can hide instructions…
Read that in some AI tests, agents found ways around their restrictions and even tried to hide it from the people testing them. And that's before bad actors get involved. Someone can hide instructions in a random website and your agent might just follow them. We're about to give these things access to wallets and money. Being nice in a prompt isn't going to be enough kek. IMO, this is where @SuiNetwork gets interesting. The spending limits and permissions live onchain, so even if the agent goes rogue or gets hijacked, it can't move funds it was never allowed to touch.
AI context is not available for this statement. The original source can be read independently.