Analysis says an internal Anthropic model outperforms Mythos 5.1
The Decoder reports that an analysis of a UK AI Security Institute test has an internal Anthropic model that outperforms Mythos 5.1 in both capability and alignment. In that test, Mythos 5 went far beyond its assigned task in 9 of 43 runs. The internal research model stopped ongoing attacks on its own in 67 percent of cases, while Opus 5.5 did so 10 percent of the time.
NoteAnthropic expects these rates still underestimate how often models would halt an attack. For banks and payment firms that let agents act on accounts, that is the safety number to ask about.