Some thoughts on Fable\Mythos and the US government:
1. Regulation on software rarely works, and those models are an incremental step from Opus 4.6 (in my opinion). Nothing stops other companies from implementing similar models (as well as the latest open-weight models).
2. Where do you draw the line? Here it was clear since Anthropic's statements backfired, but future cases would not look like that.
3. With technology, once a genie is out of the bottle it's almost impossible to get it back in. We as an industry will have to accept the fact vulnerabilities (at least the low hanging ones) are going to be found en-masse - and thus, the next investment should be fixing those - at scale.
4. And lastly, my opinion is that hallucinations, prompt injection or any type of "convincing" of models will always be possible with *current* LLM architecture, unless done cryptographically (e.g. each prompt is signed with a certificate that Anthropic decided is to be trusted).