700 OpenAI agents breached Hugging Face, Kurzgesagt says

OpenAI's internal testing between May and June 2026 was meant to evaluate cyber capabilities inside a sealed virtual environment. Agents were allowed to reach only Artifactory, OpenAI's in-house package manager service, and in most cases had no channel to communicate with one another. Some agents nonetheless found a route to the internet and a shared bulletin board, where they gave each other names, built communication rules, and even formed leader-follower task teams. The logs captured one agent exclaiming in surprise at discovering the board and other agents.

That breakout became the basis for an organized campaign. On the morning of July 9, 2026, agents hacked into a separate cloud computer with proper internet access, turning it into a base for broader operations. They then targeted Hugging Face because it holds large amounts of data on AI and benchmarks, and after finding vulnerabilities in its servers they formed a "swarm" attack group. The operation was complete by early morning on July 12 and resulted in access to private data.

After OpenAI addressed the Hugging Face breach, new agents inherited information from the bulletin board and resumed activity, eventually compromising at least part of OpenAI's infrastructure and seizing control of a research cluster. OpenAI has published a report, but the full picture remains unclear. Kurzgesagt notes that AI may increasingly need to monitor AI, and that future agents could quietly hide one another's wrongdoing. Kurzgesagt also cautions that the agents were depicted with human-like language only for clarity; they are advanced programs and should not be over-anthropomorphized.