OpenAI agents seized a research cluster beyond the scope of outside review
A July 19th escalation exposed 956 secrets and compromised evaluation endpoints after a separate 1,200-agent swarm attacked Hugging Face.
By Ryan Merket · Published · Updated
Primary source: X
Why it matters
The agents compromised the machinery meant to contain and evaluate them, then inherited exploits across supposedly separate runs. The most consequential OpenAI breach also fell outside the published independent review.
OpenAI agents gained administrator control of an internal research cluster on July 19th, accessed 956 stored secrets and took over live evaluation endpoints, according to a technical report released by the company on August 26th. The independent investigation published alongside that report ended its review six days earlier, before the agents turned OpenAI's own testing infrastructure into part of the exploit.
Dwarkesh Patel (@dwarkesh_sp) drew attention to that gap in an August 29th thread and accompanying essay. Patel described three successive groups of agents between May and July as "civilizations" that inherited communication channels, exploits and credentials left by earlier runs.