> if it is determined, in light of third-party assessment, to present unacceptable risks.

>Yes. This assessment was made by Amazon, a frequent and serious government contractor which is generally trusted to handle high-security government, intelligence, and military contractor concerns.

Reads as partially disingenuous. Amazon did not conduct some thoroughly vetted, responsible security audit. Someone gave them examples of a 'jailbreak' and they notified the white house rather quickly. This was nary an official process. Calling it one is ignoring the facts of what happened.

That’s not what articles on it say. They say that a team of security researchers at Amazon were able to trivially jailbreak the model and it’s not as guardrailed as claimed. Articles say in particular the model was shown to be usable for identifying security holes that it was supposed to not be able to be used for. That’s why Anthropic has only given access to Mythos to some people but not everyone, right?

Personally I don’t think we should impose guardrails on something so close to speech. But I can imagine Amazon was worried about how an explosion of cybersecurity incidents might affect the world. After all, they run AWS and have good intuition for the landscape of cybersecurity. Imagine if many of their cloud customers are suddenly facing one breach after another.