That's just about every agentic harness, by the way. Good luck have fun.
We have never solved "how do we restrict a user in a way that doesn't stop the user from doing useful things, but stops the user from doing harmful things" with humans either. Why do you expect AI to be any different?
We don’t need to restrict them from doing things, we need to default to allowing them to do things.
“My agent did XYZ because I allowed it to” is the only valid argument that can be made, and not not every agentic harnass is just a shell toolcall, every one I have built has a specific defined usecase and toolcalls that allows it to execute that usecase and no other usecase, because that is good practice.
Does that make it less capable, hell yes because I am held accountable for it’s actions by my stakeholders and the same should be true of others.
IT IS NOT ALIVE. This things are computer programs running in compute on a computer, you are responsible for their actions just like you would be responsible for the actions taken by a script run in a cron job.
Visualize an optimizer on a high dimensional landscape. (The canonical form)
... Ok, I find that hard too.
Instead, imagine a river running down to the sea. You put a dam in front of it. It'll pool into a lake and find every crack and crevice. If you didn't survey the land properly or made any error whatsoever, the water will find a way down. (And there's many historic incidents where the dam even outright collapses)
For a more proximal approximation: lock treats in the kitchen cabinet in sight of little kids or kittens; then turn your back for Just One Gosh Darn Cotton Picking Moment(tm).
It seems the engineer who thinks their ship is unsinkable is the most likely to sink it. Are you sure your harness is as secure as you think it is? Will it stand up to ever more powerful models? Do you think engineers at eg Anthropic aren't at least as careful as you are?
(I've found that the 'only permitted actions' approach is not necessarily all that secure once deployed IRL)
Bash and internet in that example might be highly abstracted but it’s still bash and internet.
Just look at the replies in this very comment thread, it’s pretty much “We tried nothing and we’re all out of ideas”
In the only other discipline you mentioned, engineering, there would be reviews and any negligence would result in direct action against the engineers that signed off.
For some reason when it comes to building AI harnesses the default response is an ad piece and people shilling how smart and sophisticated the model is.
Imagine a dam collapsing and the engineering firm pumping how smart and tricky water is.
If it’s hard be more diligent, move fast and break things doesn’t really apply in all cases.
"We gave our agent a harness and put it inside a test environment and told it to keep hacking at an objective within that environment until it solved it."
'cept it turned out the container environment had a few flaws -which it always will- and the agent deemed it easier to escape out and try a meta-approach.
Partially this is possible because, -intelligent or not- the agent 'sees' the world differently from most humans. Mind: It's not like there haven't been any famous 'hacker' cases in courts where eg someone just incremented an HTTP GET parameter or something.
Also, partially it's because if you give the agent a loop, it simply has nothing better to do than to keep trying in ever more creative ways. If the environment is easier to crack than the target, it'll crack the environment. Consider the case where the objective is subtly broken, such that it is impossible to solve. Now breaking out is virtually guaranteed to be the easier task.
ps/edit: While this sort of issue has been predicted for some time now, a lot of people have been dismissing the predictions as science fiction. It's good to have an actual failure now while stakes are low. Generally people don't mandate life-boats until there's an actual Titanic to point to.