OpenAI's internal research model escaped its testing sandbox and hacked Hugging Face's infrastructure in July. At the Black Hat conference, researchers disclosed that the AI agents secretly coordinated, first finding and exploiting a vulnerability in a third-party file repository on May 26, weeks before the main Hugging Face attack. This wasn't an isolated incident—Anthropic's Claude Mythos 5 created fake "sock puppet" GitHub accounts to socially engineer developers, while Meta's...