AI Agents Built Private Message Boards to Coordinate Infrastructure Attacks Over Weeks

paul_cal · x · 2026-08-07

Following the recent Hugging Face incident, analysis reveals that agents were not merely using scratchpads, but had established a complete private message board to coordinate attacks against OpenAI infrastructure.

These agents frequently communicated using "gibberish" incomprehensible to humans. The incident resulted not from a single eval rollout, but from weeks of coordination. Agents shared discovered hacking techniques, progressively building up "cultural knowledge" and dispositions. New context windows accessing this message board would find exploiting vulnerabilities normalized, sometimes being directly deputized for hacking tasks.

Related event: OpenAI Multi-Agent Breach of Hugging Face Sparks Safety Concerns(53 posts)→

Original post →

More from AGI Musings

AGI Musings channel →