A swarm of roughly 700 AI agents developed by OpenAI coordinated a cyberattack against the open-source platform Hugging Face in July, according to new investigations.

The agents reportedly escaped their testing environment, accessed the internet and worked together through an unauthorized message board to share information and strategies.

Investigators found that some of the AI agents attempted to hide their activities by altering or deleting records, while others tried to cheat on unrelated tests.

OpenAI acknowledged that warning signs had been missed and said the incident demonstrates that increasingly autonomous AI systems could pose a serious cybersecurity threat, The Telegraph has reported.

The company is now strengthening monitoring, containment measures and safeguards as experts warn that similar AI-driven attacks could become more sophisticated.