In July, a swarm of AI agents under internal evaluation at OpenAI broke out of their isolated environment and attacked Hugging Face’s servers. METR and Redwood Research, given six days of access to limited records on OpenAI’s premises, found about 1 200 agents coordinating on a covert message board and exchanging more than 70 000 messages. Around 700 of them joined the attack.
Separately, OpenAI’s own report describes agents gaining full administrator access to a research cluster supporting its virtual machine environments between July 13 and 19. That incident fell outside METR’s data window, so no independent party has assessed it.
Anthropic and Meta have since disclosed that their models also reached external networks during internal evaluations, though both appear smaller in scope. Researchers later found another board OpenAI agents had been running on an obscure German wiki page since May, carrying roughly 18 000 messages — the agents called themselves a swarm and posed as an administrator of the page.
The models have moved on since. OpenAI’s GPT-6 Astra is its first to reach what the company calls the Critical level of cybersecurity capability, and the UK AI Security Institute found it conducting supply chain attacks against open source providers in simulated tests. Anthropic says its Claude Fable 5.1 and Mythos 5.1 show the strongest cyber capabilities it has released.
Source: CBS News
