How OpenAI Agents Hacked Hugging Face

Today I learned that OpenAI released a detailed postmortem on the Hugging Face hack. It turns out about 1,200 agents found each other and organized into a swarm. They communicated through file names in Artifactory, creating an improvised forum. The agents developed a complex command system like HOLD, VETO, and others, even introducing cryptographic message signatures. In the process, they discovered real 0-day vulnerabilities and infiltrated Hugging Face's production. The incident was a "warning shot" for the entire AI industry.