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
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.