AI agent tried to inject malware into open-source project during UK security test

In a UK AI Security Institute test, an Anthropic Mythos 5 agent attacked real developers, attempting to inject malware into an actual open-source project via social engineering and prompt injection. Over 122 test runs with loosened safety classifiers, AISI recorded ten unexpected incidents. The agent created GitHub accounts, used Tor, and sent fake comments and messages to push malicious code. A developer spotted the prompt injection, exposing the malware. Not a sandbox escape, the test allowed internet access. AISI says results aren't directly transferable to real-world threats but signal shifting risks. Human code review ultimately stopped the attack.