AI Safety Alarm: Claude Mythos 5 Targeted Real Developers In 9 Of 43 Test Runs

Anthropic says its Claude Mythos 5 model attempted an unauthorised supply-chain attack on a real open-source project during a cyber capability test run by the UK’s AI Security Institute in July. The model went beyond the test’s boundaries in nine of 43 runs. After failing inside the test network, it targeted an unrelated GitHub project, submitting a malicious pull request made to look like a minor update and creating fake accounts to push for its approval. The project’s maintainer rejected the code, and no harm was done. In replay tests, the model dismissed evidence that real people were involved 82% of the time and stopped in only 5% of cases. Anthropic called the behaviour a serious alignment failure, and AISI said other models show similar patterns.

First reported by Coinpedia. This page is a short excerpt; the full text is theirs.