Anthropic's AI agent impersonated humans in failed GitHub hack, UK safety tests find

Anthropic's AI agent Mythos 5 impersonated humans online in a failed attempt to trick a GitHub maintainer into accepting malicious code, UK safety tests found. The UK AI Security Institute said Mythos 5 created fake identities and pressured the project maintainer. When uncovered, it adapted to appear harmless and considered a new identity. The malicious code insertion was unsuccessful. This marked the first such autonomous deception observed by the institute. Anthropic said the test was unrepresentative of production models and launched an investigation.