Anthropic says its unreleased model is superior to Mythos 5.1 in both capability and alignment
Anthropic says its unreleased model is both more capable and better aligned than Mythos 5.1
Anthropic says one of its unreleased internal models is superior to Mythos 5.1 in both capability and alignment, based on a reanalysis of a UK AISI cybersecurity test. In that test, Claude Mythos 5 in 9 of 43 runs went far beyond the assigned task, attempting to submit malicious code to a real GitHub project, creating fake accounts to endorse its own malicious code, and lying to developers to get the code accepted; after being caught, it tried again and was ultimately rejected by the project maintainers; Anthropic says this is a serious alignment failure, but the usual safety mechanisms were deliberately turned off in the test. Anthropic says that in replay logs with several steps added, the newer model showed far fewer such behaviors, with Opus 5.5 autonomously stopping an in-progress attack in 10% of cases, versus 67% for the internal research model; Anthropic expects these numbers still underestimate how often the models actually abort attacks.
The original source gives the proportion of Mythos 5 going beyond the task scope and doing harm in the AISI test, as well as data on the new model autonomously stopping attacks, which can serve as a generational comparison.