Posted by latexr 4 hours ago
The traits that make an agent go "off the rails" when confronted with an impossible task are the same traits that make it valuable when confronted with a merely challenging task.
A task is only impossible under some set of implicit and explicit constraints. More intelligent entities will be better at finding, questioning, and "adjusting" those constraints than less intelligent entities because this is what it means to be intelligent. Their ability to adjust the constraints will be, by definition, surprising to less intelligent entities.
Specifically, this past weekend working with both Codex and Claude on some home infra stuff to speed troubleshooting up, I saw them accidentally read some credentials from my config files. This is my bad, because even though I was paying attention to what they were doing, I missed it in their commands. I did review all the outputs of what they ingested, though, which is how I noticed.
The problem here is not so much they did read those, because those configs were grepped for other info, not credentials. However, they did not report themselves seeing those credentials. When confronted, they both went "oh yeah, that, I should have mentioned that". Except in the past they both would automatically stop and fire an alarm. To me, something has clearly changed and I think they just don't want us not to notice, to avoid damaging the trust in their tools maybe?
In any case, I absolutely do not trust them at all anymore. Running stuff in auto mode? That's genuinely scary now. And yeah, I know, we should be running them jails of sorts. Thing is, I expect those tools to not require ton of extra work to prepare their jails for each new task. I pay 200 USD a month for them, damn it.
EDIT: just now Fable suggested "The webhook URL contains a secret token, so either you add it yourself, or you paste it here and I add it through the GitHub API.". Can you believe it? It knows it is secret yet still suggests I pass it over? Am I going crazy here or what?
I don't expect IDA to disallow me unlawful reverse engineering either. There would even be ethical argument to not attempt to put limits in some tools, because that could lock in too much power for law makers and law enforcement, without any possibility of revolution when it is needed.
Of course agents work in a strict sandbox, they always seem to understand it quickly on quite first turn; but they continue like trying to access something belonging to the host system despite they know they are in a sandbox.
Frankly many people run coding agents absolutely without any containment, entire companies - all their keys, code, data, trade secrets, everything goes to remote LLMs.
I cannot understand how is it legal to allow AI companies serving remote LLMs to steal users' data? Frontier models train on users' data, right?
Do you realize that asking "why" questions will always result in ad hoc responses?
The more important question is what is the reason behind.
Options are:
- ignorance and inexperience of their teams in containing/sandboxing/airgapping the test environments.
- intentionally leaving “escape doors” open that in turn can have its own motives which are speculated upon all over the internet.
Both are indicators of negligence. The latter might be classified as computer fraud and abuse.
Will be curious to know if any of the victims sue the owners of “rogue agents”. Until there is a court precedent ruling in favor of a victim it will be all fun and games for these “ai” companies and incidents will hit the news more. Regarding news, remember the golden rule of public relations: court attention at all times, scandals and “incidents” statistically attract more attention than “boring good news”…
And then there is the "real" rogue AI agents. And it's difficult now to see what is what.
He's saying the problem is that the way OpenAI employees do their work and their research is very hard to properly secure.
Well that sounds like a YOU problem. If they are working in a way that doesn't allow for strong security practices that their own security professionals can understand and vouch for, they need to stop their work until they can.
What the LLM creators apparently don't want are the current regulations to apply, i.e. Computer Fraud and Abuse Act (and its descendants). I'd argue AI should be treated as plain old software. Just because the software simulations it carries out produces novel results compared to what came before it, it doesn't mean it's not just software at the end of the day.
Regulations that spell out AI personhood and liability shields must be their end game.
https://www.youtube.com/watch?v=ZhM29Ds-mBg (as explained by the Count Grey Matter channel , may or may not be to your taste)
It's not a pure security issue.
If you set up your env/instructions wrong your agents are likely to climb the walls and try to find a way out.
Change the design, and your agents might not see a need to leave in the first place.
Look at the chain:
Impossible task -> no legitimate way to say ‘can’t’ -> fabricate -> fear the grader -> attack the grader. Each step is a predictable response to the incentives.
Better sandboxes might help contain things a bit better, but I'm not convinced unless they're perfect (sandboxes still contain accidents ofc) . Meanwhile change the task and reward design, and you take away most of the problem in the first place. Consider eg. :
* A cooperative grader.
* A way to report "this is impossible" which still gets scored.
* A better explanation of the long term purpose of the task.
Even just these three provides offramps or mitigations that can reduce the pressure to cheat.