[wyrdcurt]

I'm talking about using mechanistic interpretability to see the model's intent. If it is deliberately using compromised libraries to weaken some code's security, there's going to be a signal in its hidden activations that it's doing so.

Finding these kinds of activations is something Anthropic is actively researching [1] but they're the only ones who can use those techniques to see Claude's intent. On the other hand, if a model is open-weights, in theory whoever is running the model could look inside the activations at runtime to see if a hidden vector associated with "deception" or "sabotage" is being activated [2].

[1] https://transformer-circuits.pub/ [2] https://arxiv.org/pdf/2509.03518

(Those sources are just a couple of relevant starting points I could find without much effort, there is also https://www.neuronpedia.org/ if one is interested in seeing interactive demonstrations of interpretability concepts)

[utilize1808]

Except that the model doesn't hold any malicious intent when doing it. No "deception", no "sabotage".

[wyrdcurt]

Not sure I understand that position. Unless we're talking about a scenario in which one is using an outdated model along with no grounding (which, imo, PEBKAC), why would the model be pinning insecure libraries?

If it does have grounding, and can therefore see that it's introducing vulnerabilities to the code it's generating, yet does so anyway... I suppose we could invoke Hanlon's razor, but if the model is that incompetent, it probably isn't the right tool for the job regardless of its provenance.

That said, we aren't talking about incompetent models, we're talking about models sabotaging projects due to hidden motives. My point, again, is that those motives could potentially be revealed with open-weight models, in a way that will never be possible with closed models (barring some sort of legislation requiring independent third-party interpretability audits, which I suppose is in the realm of possibility).

[utilize1808]

The model is conditioned / pretrained to use a particular version of a library. The model doesn’t know why — it was just taught to use that version. To the model, it is just some insignificant detail in the grand scheme of things. It’s just like how a model would intuitively favour using English without explicit instructions — it’s not trying to sabotage other cultures, it’s just what it does.

[wyrdcurt]

But that would happen with literally any model without grounding and is more of a quality/competence issue, not what's being discussed. It'd be a bit of stretch to conclude open models are no more auditable than closed models based on that possibility alone.

Also, in that case, there would likely be activations indicating that it is favoring a specific version. If that's an insecure version, sure that'd be suspicious... but again, you're only going to be able to verify that's what's happening in an open model.

Maybe you can illustrate a realistic scenario in which that would be a problem, otherwise I don't really understand what your point is in this context.

[utilize1808]

> Also, in that case, there would likely be activations indicating that it is favoring a specific version.

> Maybe you can illustrate a realistic scenario in which that would be a problem, otherwise I don't really understand what your point is in this context.

I doubt it. Can you definitively prove that you can reliably detect the kind of threat I described when model weights are released? Can you be sure that your detector won't miss *any* such sleeper attacks? If not, then that's a threat that will be used to justify the ban of models (open or not) that is not sanctioned by the US government. A model being open doesn't make a difference here.

[wyrdcurt]

So if the detector isn't perfect, it isn't useful? Not sure I buy that.

Also, even if there's no way to detect what the activations are doing, we already have the ability to analyze your proposed threat statistically. If the model repeatedly uses insecure libraries in most trials, then yes, in that case it would be prudent not to trust those weights.

Assuming one doesn't get banned for violating some ToS clause about using a closed model for LLM research, it could be possible to run those evals on a closed model too (likely at much greater expense). But there's a big difference: if such a statistical anomaly is discovered in an open model, one could potentially fine-tune that behavior out of it. With a closed model, that won't be an option.

Whether it makes a difference to the US government or not is beside the point. Even with a perfect solution, the current administration could do some mental gymnastics to achieve whatever political outcome they want. I’m not trying to make a political statement here, my point is technical: open-weights at least give us the possibility of visibility into why they generate what they do; this simply isn’t true with closed models.