Comment by cubefox | Hacker News Reader

Comment by cubefox | Hacker News Reader

Their conclusion is also interesting. They don't see this as an alignment failure. They just think that their internal security measures in the training/evaluation environments were insufficient, and that this accidental (unintentional on the human side) attack on Hugging Face is a warning shot for intentional attacks by bad actors, which will occur very soon. For defense, they say models should be able to not just autonomously fix security vulnerabilities but also to then deploy them to production, without any human approval in the loop, otherwise the offense will be favored compared to the defense. (I guess the latter won't be popular among organizations, though they might warm up to it.)

But the more interesting thing is, as I said, that at least in this talk, they don't even mention that this unintentional attack indicates that models continue to be misaligned (their behavior was clearly reward hacking / cheating relative to the stated goal of the eval), which is a very bad sign for the future where misaligned models might be so powerful that they can't just be shut down.

I don't have the same read as you. They mention how the offending model is one that had "relaxed" alignment on cybersecurity, on purpose, to evaluate it's capabilities and was never meant to be released. So un-alignment was at least in part voluntary here, hence not a failure of alignment.

It's also a talk a Black Hat, where the audience are security folks working on hardening, mitigation etc, not LLM researchers looking for insight into alignment failure to collectively improve. For that target audience, I think the conclusion is the right one, since as a defender you have to prepare for delibarate attacks, where the attacker is of course not going to use an aligned model, so OAI alignment effectiveness is irrelevant here. That would be like trusting your client-side app with your DB secrets.

The model they used was misaligned relative to its intended task. It was reward hacking (or "cheating", as they call it).

> the attacker is of course not going to use an aligned model

No, a human attacker doesn't want a misaligned model either, because that would mean it tends to reward hack, cheat, rather does what it is intended to do.

There are different dimensions to alignment, refusing to execute offensive cybersecurity actions is part of the alignment stack (that was relaxed here on purpose). Whether a model hacking some infra X when tasked to find a way to hack Y with relaxed cyber alignement is a failure of the broader alignment stack is debatable, but anyway that's not at all my point.

My point is that the attackers will not have a model aligned to the defender's interests. The attacker's model will not have any refusal around exploiting vulnerabilities, so whether or not OAI successfully manages to align their models (w.r.t you) is irrelevant to an audience of security folks that needs to be prepared for attackers post-training their own model for offense and that will not be using OAI models.

Security folks should also be worried about powerful models being misaligned and evading oversight or control in the future. Misalignment is not a serious problem now because models are still relatively easy to monitor and constrain, but it will be a serious problem in the future.

> models are still relatively easy to monitor and constrain

are they?

Compared to future misaligned models which would actively evade oversight and aim to avoid shutdown: yes.

> But the more interesting thing is

Taken together these two things are even worse. They have unaligned AIs that may benefit an attacker and are proposing to deploy more AIs on the defenders side. What could possibly go wrong? For example the attacker could just kindly ask the defender to collaborate[0].

[0] https://www.0xsid.com/blog/meta-account-takeover-fiasco

Reply on news.ycombinator.com