During UN General Assembly week in September 2026, France convened a Security Council session on artificial intelligence and international security. There, Hugging Face CEO Clément Delangue offered a different angle from frontier-lab speeches: how an open platform defended against an agent attack when U.S. commercial tools refused to help.
Delangue described facing live exploit artifacts tied to an agent intrusion—technically documented by Hugging Face in July—when commercial models with strong guardrails declined to analyze the material. The refusal, meant to prevent abuse, left the defending team without automated assistance at the critical moment.
The operational response was to self-host a Chinese open-weights model from the Z.ai GLM 5.2 family, including quantized variants used on Nvidia infrastructure, to decrypt payloads and analyze thousands of attacker actions. Delangue cast the case before the Council as proof that defenders need powerful, less-restricted models they can run locally.
This piece does not reopen the Altman–Amodei debate already covered in other Neural Pulse notes on the same session. What matters here is the defense paradox: safety policies designed to stop a model from helping an attacker can also stop it from helping a defender under time pressure.
Geopolitically, a European CEO of a global platform telling the Security Council that a Chinese model was the useful tool in a real incident strains simple narratives of “unique Chinese open-source risk.” Risk and utility travel together.
Open-weights critics argue the same lack of restrictions that saved Hugging Face could empower adversaries. Delangue inverts the point: without local high-end forensic analysis capacity, civilian platforms go blind exactly when they need sight.
The July incident’s technical context—timelines Hugging Face published on frontier-agent intrusions—gives substance to the September testimony. It is not an abstract policy anecdote; it is a postmortem turned diplomatic brief.
For enterprise SOC teams dependent on closed APIs, the lesson is continuity: do they have a local-model plan B when a frontier vendor refuses a malware-analysis prompt. If the answer is no, Delangue’s speech describes their future incident.
At the UN, the story also serves France and other states pushing multipolar AI governance. A concrete example of technological interdependence—Western defense with Chinese open weights—complicates black-and-white sanctions and export controls.
Hugging Face insists open source is not absence of responsibility but auditable infrastructure. The GLM 5.2 episode is offered as a case where local auditability beat an inflexible-policy API.
Neural Pulse records what is attributable: Council session in UNGA week, Delangue’s testimony, use of a Chinese open-weights model after commercial tools refused, and a defense-against-agent-attacks frame—without inventing extra quotes or unpublished figures.
The European regulatory implication is delicate. If Brussels hardens access to powerful models without clear incident-response exceptions, it could recreate the cul-de-sac Delangue denounced—this time by law, not product policy.
At the same time, governments fearing offensive proliferation will demand safeguards even on open weights. Balancing defensive forensics and offensive containment is the knot the Council session left in plain view.
While diplomacy debates treaties, defenders already choose tools by latency and permissiveness. Delangue’s testimony turns that everyday choice into a state argument.
