This is a special post for quick takes by DW11. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Inspired by this Jack Clark tweet (as well as the emphasis on external review in RSP v3):
https://x.com/jackclarkSF/status/2053847777889964376
Has Anthropic made Mythos available to any external AI safety organisations for internal use (not just evaluation)?
There do seem to be some NDAs relating to Mythos given the lack of people on twitter talking about having access, so perhaps they are using it and are just keeping quiet. But it seems reasonable that they should be given access, given that it seems to be fairly widely-deployed for cyber at this point (not to mention internally at Anthropic).
Given all the discussion of Mythos, I'm surprised that I haven't seen anyone point out that OpenAI launched Trusted Access for Cyber alongside codex-5.3:
https://openai.com/index/trusted-access-for-cyber/
It seems that if you try to use 5.3 for cyber without verifying your identity, you get routed to 5.2:
https://developers.openai.com/codex/concepts/cyber-safety
So the practice of gating capabilities based on cyber risk is not totally brand new. I'm curious to see how they handle this for 5.5, given that I expect it would be hard to generally release a model anywhere close to Mythos now without causing backlash.