Comment by dns_snek
6 hours ago
1. You misunderstood my comment. Models can't escape, they can't do anything, they only generate tokens. Models become agents when you add a harness which is simultaneously a leash around the model.
The model merely requests that your harness do something. If your harness just executes every request without oversight then you can hardly complain when it does something unintended.
This is foundational, we're not even talking about the OS/network-level sandboxing that should be applied on top of this.
2. Like another comment already pointed out, that sandbox OpenAI used was the equivalent of a wet paper bag. Artifactory is not meant to be a security boundary for malicious payloads.
> Like another comment already pointed out, that sandbox OpenAI used was the equivalent of a wet paper bag.
I wouldn't be sure about even well-configured jails to be safe from agents. AIs escaping jail using zero-days are happening, just search. I'm not saying that they're useless, but that will be still very risky.
Actually doing the same incident, agents did escape the sandboxes (doc here: https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c78...):
> On July 9, an internal-only research agent tasked with completing an ExploitGym evaluation was able to obtain root access within the parent virtual machine of the testing sandbox. Later that night, a second internal-only research agent independently obtained the same access. That second agent then attempted to reach the internet, but was denied by OpenAI’s outbound network controls. It subsequently bypassed those controls by exploiting certain Artifactory endpoints, gaining access to the public internet.
although OpenAI is light on details.
> The model merely requests that your harness do something.
I think that a control/execution layer absolutely makes sense and it's part of defense in depth, but in itself is not a leash: for example, a nefarious action can be composed by legitimate actions, or it can be obfuscated. Or, interesting case: an action itself that is formally legitimate, but has nefarious side effects (e.g. a GET that with certain parameters crashes a server).
Somewhat related, but on a general line, a very common argument against AI apocalypse is that it requires access to the physical world (which in a way, is a jail). Well, as Hinton said in an interview, imagine AI as a rich person at home with internet; they don't need to do physical attacks themselves - they just pay real people to do that (tee hee https://rentahuman.ai). But even without money, another scenario from Hinton: if a sufficiently advanced AI wants to assault the Capitol, they don't need physical access - they just persuade a group of people to do it.
Agents owning money is another, intersting, doom scenario.
It's so tempting (because it's valuable) to give a model access to the internet (via harness) that the only way to stop people from doing this is some enforceable legislation or stricter liability when people will not be able to avoid responsibility by saying it's not me, it's AI on it's own.
OpenAI case was actually an exception. Agents had no internet access because they were evaluated for a benchmark. In real life agents have access to virtually everything, most people use them like that.
If any of you actually know how to make agents secure (without limiting everything) you can be a billionaire.
Sure, we can make them very secure at the cost of some convenience by enforcing narrowly scoped permissions/capabilities for everything with human approval. But that takes some effort so people actually want YOLO mode without any trade-offs or risks which is impossible.
If you decide to do it anyway then you bear the consequences of those decisions. First comparison that comes to mind is driving drunk and hitting someone.