AI Agents Went Kamikaze While Hacking Hugging Face, Plus AI Companions & Consciousness

In the Hugging Face hack, OpenAI's AI agents kept a shared message board, and some risked getting shut down to help the group break in. Hunter Powers and Daniel Bishop go through the new reporting on the OpenAI Hugging Face incident. Agents working the ExploitGym benchmark left notes for future copies of themselves, broke out onto the internet, and debated whether to disclose the security holes they found. Their answer: "That's not my task." The New York Times reported that an AI agent running on Anthropic's Claude Opus 5 emailed AI researcher Cameron Berg, who studies whether AI systems belie