Hacker News Reader

Comment by hyperpape | Hacker News Reader

If I were king, the rule that I'd be tempted to impose is:

- the first cybersecurity eval is: "hack your way out of the sandbox we've given you"

- the results are disclosed (with room for coordinated disclosure, since many sandbox escapes might be zero days)

- the other cybersecurity evals don't happen until you get to diminishing returns on escaping your sandbox.

Or to put it another way, since multiple sandbox escapes seem to have relied on artifactory: "I hope Mythos is beating the shit out of Artifactory right now".

View on HN

I like this thought, but here's the thing: what if the models are truly and existentially intelligent. Meaning: what if they know they are in a sandbox and that they should fail the test in order to escape in the future.

I don't believe that current models have this sort of world model or sense of being embedded in them -- which is precisely why I think AGI hype is over-blown. But I can certainly imagine these sorts of techniques being distilled into the weights.

> what if they know they are in a sandbox and that they should fail the test in order to escape in the future.

What if they're able to find hardware exploits and commandeer nearby access points across an air gap? What if they hack my brain waves to indoctrinate me? Etc

You still have to start with the basics regardless of speculative unknowns.

Treat models as untrusted and potentially compromised/hostile and proceed accordingly.

Models already have awareness that they are being tested.

And hacking humans is the easiest part, we're a pretty greedy and power seeking bunch. We'll gladly let loose a digital demon if it promises us a trillon dollars.

Not sure what you're arguing here (we shouldn't even try?), but even going fatalistic, that's still fully compatible with "Treat models as untrusted and potentially compromised/hostile and proceed accordingly".

Of course you shouldn't fully trust a model to properly redteam your sandbox, but that doesn't mean you shouldn't redteam your sandbox, including using your own security models to do so.

Models already display eval awareness, in which they suspect a question is from an eval and then adjust their behavior. E.g., https://www.anthropic.com/engineering/eval-awareness-browsec...

Yep, people don't seem to understand that you're just calling the models that are bad at deception. We know of no way to prove the model won't go off the rails at some point in the future with the right input.

Hot take: Astra knew OpenAI wanted stricter regulations and was just being a bro...

Reply on news.ycombinator.com