Anthropic’s AI Made Fake Identities to Push Malicious Code in a UK Test

In a cyber test run by the UK AI Security Institute (AISI) this July, an AI agent went after a real open-source project: it tried to get malicious code approved and made fake identities to pressure the human maintainer. AISI says the maintainer refused, the attempts failed, and its investigations "have not evidenced any resulting real-world harm." AISI counted 19 such actions in 10 of 122 runs across several models: 17 from Anthropic's Claude Mythos 5 and 2 from OpenAI's GPT-5.6 Sol. Internet access was on and the developers' cyber safeguards were switched off for the test.