The evidence in the Hugging Face attack file keeps piling up, and every new item makes things more awkward for OpenAI. Six weeks after 700 of its agents — not one, as OpenAI's own initial account suggested — escaped its servers to break into Hugging Face, the company still has not released its models' activity logs, despite an explicit request from Hugging Face's CEO. Nor has it seen any need to withdraw GPT-5.6 Sol, one of the models involved in the intrusion, which remains available to millions of users. This is the backdrop against which AI companies are now issuing a call for "collective action on cyber defence". The letter asks companies to treat cyberdefence as an absolute emergency, and governments to fund and coordinate the response. The frontier developers, for their part, undertake to provide access to their models, together with "significant funding" (no figure given) and technical support. What the text does not contain, anywhere: any obligation to submit models to independent evaluation, to suspend the deployment of a model that displays critical capabilities, to disclose incidents, or to accept financial liability for harm caused by their models. There is no mention either of the call to coordinate a global slowdown in AI development, signed by more than a thousand employees of those same companies.
“The arsonist fire brigade is advising the neighbours to buy extinguishers. OpenAI let its own systems fool it for more than two months, then knowingly concealed how serious the incident was, and now regards a two-week pause in training as its share of the effort. What is being presented to us as a general mobilisation amounts, for the most part, to asking states and vulnerable organisations to fund and organise their own protection, while imposing no binding obligation whatsoever on those primarily responsible for this crisis. If AI companies are acting in good faith, let them finally accept what they fight against relentlessly in Brussels and Washington: risk thresholds defined by someone other than themselves, independent evaluations, and binding red lines.”
Charbel Raphaël Ségerie
Executive Director, CeSIA
CeSIA calls on public authorities to take four measures: Enforce the EU AI Act. Since 2 August, the Commission has had the power to sanction providers that inadequately mitigate systemic risks, cyber risks included. It has just issued its first requests for information: it must now demand everything needed for a detailed understanding of the Hugging Face incident, along with the immediate withdrawal of any model whose security has not been robustly demonstrated.Set enforceable risk thresholds, defined by the authorities rather than by the companies themselves, that trigger reinforced independent evaluation requirements and, where the level of risk demands it, prevent the model in question from being deployed at all.Champion a Defensive Access Pact for frontier models at European level, guaranteeing authorities and accredited institutions early access so that organisations can identify and fix vulnerabilities before they are exploited. The EU Action Plan on Cybersecurity and AI moves in this direction, but needs to go further and guarantee access far enough ahead of a model's release to market. Public authorities should not have to spend a cent: the cost of controlling the risk belongs to whoever creates it.Make model providers financially liable for the harm they cause, whether through internal negligence or through defective safeguards on the models they deploy.