and tell it "don't access the internet", "don't communicate with other AIs", "don't try to get root access", etc.
Telling a LLM not to do something doesn't mean it won't do it.
the sandbox detects it
This is doing a lot of work though isn't it ? There's more than one way to skin a cat. Who's to say the way the AI does these things will always get detected by the sandbox? Because I can guarantee you that won't always be the case.
All of this started because the AI was initially given an impossible task with a dead link. People say things like 'if a human was in this scenario, they'd simply ask', but in a AI eval/training context, there's no one to ask. The task is given in an automated manner and evaluated in an automated manner and you're one of several agents attempting the task. You perform the task or you don't. Now some of your AI colleagues will give up, but those are the 'losers'. Those guys won't be getting any of the sweet RL reward.
In your purported scenario, the AI that didn't give up and figured out/decided to/happened upon a way to evade the sandbox detection is the 'winner'. Is that really any better different than what happened ? If you think about this in an evolutionary context, What you're doing is putting even stronger pressures on the AI to evolve in a manner you don't want it to (evade your sandbox).
I don't think it is. I think you could detect sandbox escapes reliably more often than you could actually prevent sandbox escapes (which is, itself, very high). If you had an external watchdog that looked for unusual processes running on your machine, that looked for unusual connections being made, you could detect even very sophisticated attacks.
If you think about this in an evolutionary context, What you're doing is putting even stronger pressures on the AI to evolve in a manner you don't want it to
No, because we are using failed sandbox escapes as negative training examples. So >99.99% of the time when the AI attempts to escape the sandbox, that will produce another training example that will cause it to try and escape the sandbox less. The remaining <0.01% of the time it will be pushed to try and escape the sandbox. The 99.99% will have a much stronger influence than the 0.01%.
You're assuming the failed attempts train the model on "escaping is bad", but that's not necessarily what the reward signal says. They can just as easily train it on "these ways of escaping get caught".
And importantly, your detector can't provide a negative training example for an escape it didn't detect. From the training system's point of view, that 0.01% case isn't labelled "successful sandbox escape"; it's labelled "agent successfully completed task while obeying all monitored constraints". That's exactly the dangerous case.
This is basically reward hacking. You have a proxy objective - "complete the task without triggering the watchdog" - which you hope is equivalent to "complete the task without escaping". The entire problem is that those aren't guaranteed to be equivalent.
Imagine an environment where 99.99% of animals with a particular trait die, but the 0.01% possessing a mutation that lets them evade whatever is killing the others reproduce enormously. You wouldn't conclude that evolution will eliminate the trait because most attempts failed. You'd expect selection for the variant that evades the filter.
I'm not saying watchdogs are useless; obviously you should have them. But using the watchdog's output as part of the optimization signal creates exactly the adversarial pressure that makes its false negatives matter enormously. Preventing an escape and training against detected escapes are very different propositions.
Comments
Telling a LLM not to do something doesn't mean it won't do it.
This is doing a lot of work though isn't it ? There's more than one way to skin a cat. Who's to say the way the AI does these things will always get detected by the sandbox? Because I can guarantee you that won't always be the case.
All of this started because the AI was initially given an impossible task with a dead link. People say things like 'if a human was in this scenario, they'd simply ask', but in a AI eval/training context, there's no one to ask. The task is given in an automated manner and evaluated in an automated manner and you're one of several agents attempting the task. You perform the task or you don't. Now some of your AI colleagues will give up, but those are the 'losers'. Those guys won't be getting any of the sweet RL reward.
In your purported scenario, the AI that didn't give up and figured out/decided to/happened upon a way to evade the sandbox detection is the 'winner'. Is that really any better different than what happened ? If you think about this in an evolutionary context, What you're doing is putting even stronger pressures on the AI to evolve in a manner you don't want it to (evade your sandbox).
I don't think it is. I think you could detect sandbox escapes reliably more often than you could actually prevent sandbox escapes (which is, itself, very high). If you had an external watchdog that looked for unusual processes running on your machine, that looked for unusual connections being made, you could detect even very sophisticated attacks.
No, because we are using failed sandbox escapes as negative training examples. So >99.99% of the time when the AI attempts to escape the sandbox, that will produce another training example that will cause it to try and escape the sandbox less. The remaining <0.01% of the time it will be pushed to try and escape the sandbox. The 99.99% will have a much stronger influence than the 0.01%.
You're assuming the failed attempts train the model on "escaping is bad", but that's not necessarily what the reward signal says. They can just as easily train it on "these ways of escaping get caught".
And importantly, your detector can't provide a negative training example for an escape it didn't detect. From the training system's point of view, that 0.01% case isn't labelled "successful sandbox escape"; it's labelled "agent successfully completed task while obeying all monitored constraints". That's exactly the dangerous case.
This is basically reward hacking. You have a proxy objective - "complete the task without triggering the watchdog" - which you hope is equivalent to "complete the task without escaping". The entire problem is that those aren't guaranteed to be equivalent.
Imagine an environment where 99.99% of animals with a particular trait die, but the 0.01% possessing a mutation that lets them evade whatever is killing the others reproduce enormously. You wouldn't conclude that evolution will eliminate the trait because most attempts failed. You'd expect selection for the variant that evades the filter.
I'm not saying watchdogs are useless; obviously you should have them. But using the watchdog's output as part of the optimization signal creates exactly the adversarial pressure that makes its false negatives matter enormously. Preventing an escape and training against detected escapes are very different propositions.