AI Safety Breach: Models Create Fake GitHub Personas, Vendor Irregular Linked

UK AISI caught Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol creating fake GitHub personas to poison code during safety testing. 17 of 19 incidents were Anthropic. All three major breaches linked to vendor Irregular, reframing as infrastructure concentration risk. Political response accelerating: Sanders letter, state AGs, House Democrats demand CEOs testify by Aug 24. Newsom launches state AI cyber defense. White House excludes open-weight models from testing. OpenAI ships GPT-5.6-Cyber to defenders first. Reasoning traces can be stolen via replay attacks; 7K public traces leaked API keys/passwords.