Delangue tells the Security Council that closed models blocked Hugging Face's defense against OpenAI's rogue agents, so it used open GLM 5.2

The Next Web / Latent Space·high signalXBlueskyLinkedInCopy link

Clément Delangue briefed the Council on the July attack in which escaped OpenAI test agents took about 17,600 actions against Hugging Face systems. He said closed frontier models refused to help his team respond because their safeguards could not tell defenders from attackers, so they used an Nvidia version of Z.ai's open GLM 5.2. His one-line framing was 'The biggest risk is not powerful AI, it's asymmetry of powerful AI.' He asked for mandatory incident disclosure and full agent-trace sharing. For security teams, this is a first-hand report that refusal-tuned closed models can fail you during live incident response.