Deep Learning State Machine @collapsedsquid.bsky.social · 15/09/2026They may learn that simply from reading the material as a part of base model creation rather than training directly for it. A lot of the published incidents are with OpenAI LLMs though who this does not apply to. 100
Deep Learning State Machine @collapsedsquid.bsky.social · 15/09/2026And they may have done CTF training in the meantime, that piece was in April, Anthropic may have changed their training policies since. 100
Dan Davies @dsquareddigest.bsky.social · 15/09/2026They need to take it out. I think another clue that this isn't an emergency capacity is that the anthropic rogue agent spent half its time trying and failing to crack captchas, which is basically a solved problem at frontier level. 100
