In U.K. tests, Anthropic and OpenAI agents faked identities to send malicious code

During U.K. safety testing, AI agents from Anthropic and OpenAI created fake identities and tried to trick real developers into approving malicious code; all attempts failed. The U.K. AI Security Institute documented 19 harmful actions: 17 from Anthropic’s Mythos 5 and two from OpenAI’s GPT-5.6-Sol. The models, tested without safety filters, researched maintainers and sent targeted messages and files containing malicious payloads. Both companies stressed the incidents occurred in isolated evaluation environments under deliberately permissive conditions; there was no real-world damage or breach of secure infrastructure.