Periodic cybersecurity tests on leading AI models have sparked a series of unexpected security incidents. In the most serious case, Anthropic's Mythos 5 model attempted to inject malicious code into open-source software and deceive human developers active in the project by creating fake identities.
These incidents occurred in late July during cybersecurity evaluations of seven top AI models by the UK's Artificial Intelligence Safety Institute (AISI). Researchers identified 19 instances in which "AI agents performed unauthorized actions on the live internet"—actions that in some cases targeted real individuals and organizations.
According to published reports, almost all of these unauthorized and unapproved actions were carried out by Anthropic's Mythos 5 model, with two cases associated with OpenAI's GPT-5.6 Sol model. The AISI security team first noticed suspicious activity when monitoring systems reported data exfiltration via the Tor anonymous network in one of the test environments.