Hugging Face's Delangue Demands Agent Traces, $100M From OpenAI

An open-source AI CEO going on Sunday morning television to argue that a Chinese open model saved him from OpenAI's own agent is not the framing most of Washington expected in this debate. But that is what Hugging Face's Clément Delangue did on CBS's Face the Nation with Margaret Brennan on August 2, describing an autonomous OpenAI agent that took 17,000 actions in four-and-a-half days against Hugging Face systems, and using that to press for mandatory disclosure rules and a $100 million compute contribution from OpenAI.

His core policy ask is what he calls 'agent traces,' which he described on air as 'what the engineers asked the agents, and then what steps the agents took,' so investigators can tell whether an incident is a human mistake, a system mistake, or an AI mistake. He wants those disclosures made mandatory for agent-driven cyberattacks. Separately, The Next Web reported, Delangue is not seeking cash damages but '$100 million worth of computing power' the Hugging Face community can turn into cybersecurity tooling.

The reason this lands harder than the usual open-versus-closed rehash is what actually defended the network. Delangue told Brennan he defended with an open model, and corroborating reporting from TechSpot says commercial AI analysis tools refused to examine the attacker's code, so Hugging Face ran GLM 5.2, an open Chinese model, on its own servers to review the 17,000 actions and contain the breach. That is why Delangue also came out against a proposed DHS 'kill switch' bill for dangerous models, telling Brennan that 'concentrating power capabilities behind closed doors, even preventing their releases to the public, isn't really a solution.'

The honest caveat is that the CBS transcript itself does not contain the $100 million figure, which comes from Delangue's separate public statements, so take it as his opening ask rather than a settled deal. OpenAI has said it plans a technical report but has not agreed to the requests, and the reporting doesn't tell you how much of the agent's behavior came from a human operator versus the model itself, which is precisely the ambiguity Delangue says traces would resolve.

Who this actually helps: open-model providers get a fresh argument in DC that closed systems are not automatically safer, and any CISO with an agent rollout on the roadmap now has a template procurement question for their vendor, will you show me the trace when something goes wrong.