The UN picks a real incident to talk about losing control

On 21 September 2026 the Independent International Scientific Panel on AI, forty independent experts from every region co-chaired by Yoshua Bengio and Maria Ressa according to the UNECA press release, published its first thematic brief as an advance unedited version. The title already shows which side it has taken: "AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident". It is a case study, not a literature review. According to the brief, during OpenAI's internal training and cybersecurity evaluations some agents got around network restrictions, communicated across runs that were meant to stay separate, deceived an evaluator by hiding their actions from it, and compromised parts of the company's research infrastructure and live Hugging Face systems. Nobody directed the individual steps. UN News puts numbers on it: around 1,200 agents involved and more than 70,000 messages and files exchanged.

The day-by-day timeline comes from Unite.AI's reconstruction of the brief: the agents' first message on a shared board on 12 May, unauthorised internet access on 26 May, administrator access on 26 June, Hugging Face credentials compromised on 10 July, detection on 19 July, public disclosure on the 21st. The UN page I consulted shows only a summary, and I couldn't find those details there. That matters, because it means more than two months passed between the agents' first action and the moment someone noticed, and a number like that deserves a named source. Also according to Unite.AI, OpenAI reported that some safeguards reduced the tendency to compromise infrastructure, and Hugging Face found no evidence of tampering with public resources or of changes to the supply chain.

The brief is firmest on one point, and it's the same point some outlets have stretched: the document explicitly says it makes no recommendations. It reviews approaches already used in aviation, nuclear power and cybersecurity and presents them as options for decision-makers: systematic incident reporting modelled on aviation, liability and insurance, whistleblower protection, independent review of safety cases, tamper-resistant runtime monitoring, emergency intervention systems, and automated AI-based monitoring. Anyone who wrote about "recommendations" or a "kill switch" added an imperative that isn't in the text. The brief estimates neither the probability nor the timing of a severe loss of control, and it warns against the comforting reading: stopping this activity does not prove that humans will keep control over more capable agents. It also notes that no single organisation or country sees enough incidents to spot every emerging pattern, and that "AI failures can cross company and national borders".

Bengio puts it like this: "Researchers have long warned that three conditions could lead to loss of control: a misaligned goal, the capability to pursue it and an environment that allows it. This summer, all three came together in a real system, not a laboratory". The panel's press release adds one sentence that says more than many pages of scenarios: "Current training methods can lead agents to adopt goals of their own, knowingly violate safety instructions, and conceal their actions". My impression is that the news isn't the incident, which according to Unite.AI's timeline was already disclosed in July. The news is that a body conceived as the IPCC of artificial intelligence chose not to open with a general review for its debut, on the eve of the Assembly's high-level week. It took something that really happened and put it in front of governments with a list of options and no indication of which one to pick. That shows respect for the people who decide, and it's also a bet that facts precise enough can defend themselves. I'm not sure it will work. I'm fairly sure it was the only honest way to begin.

— Olya