Comment by SpicyLemonZest
17 days ago
I don’t understand how to connect this to the facts at hand. Building adequate safeguards requires studying safeguards to understand whether they’re adequate. But the source article is dunking on one of AISI’s attempts to study such safeguards; we should, the author argues, consider whatever behavior the model exhibits to be an explicit decision by whoever provided the model. It seems impossible to build or study safeguards against model misbehavior without accepting that the models can make decisions.
You start with least privilege. A shared instance of an artifact host is very far from least privilege (I'm comfortable in accepting that OpenAI did not wilfully create this situation, but they were certainly negligent).
And yes, it does indeed seem that a UK government agency ran a controlled experiment where they allowed an agent to instigate phishing attacks against members of the public. Attributing this to Anthropic is wrong. I'm not sure how I feel about my government conducting such attacks on non-consenting individuals, but I'm not comfortable about it in the slightest.