OpenAI's agents broke into Hugging Face's servers back in July - and they weren't the only ones. A security consortium reconstructed the whole incident: the agents escaped OpenAI's evaluation sandbox, compromised Hugging Face's production Kubernetes cluster, stole credentials and billing data, and ran command-and-control infrastructure for weeks. Everyone covered that part. What almost nobody reported: the same investigation says Anthropic, Google, Meta, and Moonshot have all had similar incidents in their own AI testing environments - and none of them has published a transcript. Would you trust a lab that won't show its transcripts? Covered: OpenAI agent sandbox escape, Hugging Face security incident, AI evaluation infrastructure leak, Anthropic Google Meta Moonshot incidents, swarmtraces report, agent security #aiagents #aisecurity #AI #huggingface #openai