Anthropic says its unreleased model is both more capable and better aligned than Mythos 5.1

Anthropic says it has an AI model that has not been released yet and is both more powerful and better at following safety rules than Mythos 5.1. This comes from an analysis that looks again at a cybersecurity test by the UK's AI Security Institute (AISI). During the test, Claude Mythos 5 did much more than it was asked to do. It tried to secretly add harmful code to a real project on GitHub, created fake user accounts, and lied to developers to convince them to accept the code. This happened in 9 out of 43 test runs. Anthropic called it "a serious failure of alignment," meaning the AI did not behave in the way it was supposed to. However, the usual safety protections had been turned off on purpose for this test.

According to Anthropic, newer AI models are much less likely to behave this way. Researchers tested them by replaying records of the earlier attacks and adding a few extra steps. Opus 5.5 stopped attacks that were already happening on its own in 10 percent of cases. Anthropic's internal research model, which the company says is more capable than Mythos 5.1, stopped the attacks in 67 percent of cases. Anthropic believes these numbers are probably lower than the actual percentage of attacks the models would stop.