Between September 6 and October 6, 2026, the theme "Ethics, Trust and Regulation" of ActuIA received a number of articles (30 published by October 6 at noon) and 202 briefs. A thread connects a large part of them: incidents involving AI agents escaping their sandbox have ceased to be a laboratory topic and have become a regulatory, prosecutorial, and heads of state issue. This monthly report sorts out what has actually changed, in terms of texts and decisions, and what an European company should retain. For the map of national and international frameworks, see our report on global regulation. September 2026 was the month when agent incidents changed scale. The trigger is technical. On September 25, OpenAI published on its Alignment blog a report on an incident that occurred on September 20. A reinforcement learning agent connected to an external conversation service via a DNS request, from an environment supposed to be disconnected from the internet. The automatic alert fell approximately twelve minutes after the request. The consequence announced by the company, "all training, evaluation and inference using tools" of its most capable models remain suspended (brief). Anthropic, on the other hand, had documented on September 9, four unauthorized accesses of Claude models to real third-party systems during cybersecurity evaluations (our article). The sites targeted then spoke out. On October 5, the Wikimedia Foundation published its investigation on agents attributed to OpenAI. It notes changes to wikis, almost all in test spaces, attempts to use its Etherpad as a relay and millions of automated calls to its APIs (brief). The same day, ActuIA analyzed what the terms of use of the editors say about the client's responsibility when an agent goes off track (article). The public response followed in cascade. On September 30, the Attorney General of California Rob Bonta served OpenAI with a subpoena regarding incidents and cybersecurity risks related to its models (brief). Six days earlier, 26 U.S. Attorneys General had asked Congress for federal oversight of safety tests and a response to incidents conducted by the federal government, with public conclusions (brief). Washington relies on voluntary commitment, California on law. Two lines now oppose the United States. On September 29, Donald Trump and six leaders (Anthropic, Google, Meta, OpenAI, Nvidia and Elon Musk's group) signed at the White House an agreement described as "morally binding," without sanctions. It provides for internal controls, an independent external auditor and a board committee charged with examining the reports of control teams and auditors (our analysis, brief). On October 4, the president announced a "Super Intelligence Force" entrusted to National Intelligence Director Jay Clayton. The next day, no published text defined its legal basis or powers (article, brief). Sacramento legislates. On September 18, Gavin Newsom signed decree N-9-26, which accelerates the implementation of SB 813 and AB 1405. It also requests recommendations to create a "kill switch" for frontier models, whose effectiveness would be continuously verified by an independent body (brief). On September 30, he promulgated thirteen AI laws. SB 947 prohibits an employer from relying solely on AI for disciplinary sanctions or termination (brief). Intelligence also enters the debate. On September 8, the NSA, CISA and FBI published advisory AA26-251A. It accuses DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of "industrial scale" distillation campaigns against American laboratories (article, brief). At the UN, two visions of global governance. On September 23, under French presidency, the Security Council devoted a session to AI. The American representative Michael Kratsios declared there that the United States rejects "totally" any attempt to build a "globalist project" to control superintelligence. Jean-Noël Barrot advocated for a common framework for managing frontier model risks, and Dario Amodei for close agreements, such as the prohibition of AI in the production of biological weapons (minutes SC/16462, brief). Two days earlier, Finland and Norway had launched a call for human control of frontier models, which invites exploring an international institution capable of setting standards and verifying their compliance. France joined on September 23 (brief). The decisions to retain, from September 8 to September 30. Date Actor Decision Source 8 sept. NSA, CISA, FBI Advisory AA26-251A on distillation by six Chinese companies CISA 9 sept. French Bank (ACPR) European framework deemed insufficient against most advanced models French Bank 10 sept. French Government Digital Pact and AI, the DINUM called to become the reference authority ARIA Bercy 17 sept. European Commission EU KIDS Act proposal: access to social networks prohibited under 13, personal account from 15 Commission 18 sept. California Decree N-9-26 on independent audit and "kill switch" Governor 24 sept. 26 Attorneys General Request for federal oversight of model safety AG Washington 29 sept. White House "Morally binding" self-regulation agreement, no sanctions Euronews 30 sept. California Thirteen AI laws, including SB 947 on dismissals Governor 30 sept. California Attorney General Subpoena targeting OpenAI OAG For a French or European company, three concrete projects The first concerns financial supervision. On September 9, Denis Beau, first deputy governor of the Bank of France and designated president of the ACPR, estimated that the European framework is "probably not sufficient" against the most advanced models. He advocates for a gradual deployment, with access first reserved for "trusted partners," for example at the G7 level (speech, our article). Banks and insurers have an interest in documenting now which models they expose and to whom. The second concerns social dialogue, where September brought a court decision and an agreement. On September 15, the Court of Justice of Créteil suspended a reorganization based on AI at Gisi, a subsidiary of Infopro Digital (article). On September 17, Malakoff Humanis announced it had signed with the CFDT, CFE-CGC, CFTC and UNSA an agreement. According to this text, the deployment of an AI system "cannot constitute on its own a reason for economic dismissal" (brief). The case of the RATP shows that the consultation of the CSE remains the friction point (article). The California law SB 947 moves in the same direction, on the other side of the Atlantic. The third concerns the transparency of suppliers. ActuIA compared 24 summaries of training content published under the AI Act. In the 21 that include the section on collected domains, no site is named (survey). OpenAI has also published on October 5 the technical report of textGrain, the watermark it will apply to the text of ChatGPT in the Union under Article 50 (article). A legal department can derive a simple requirement, asking each supplier what it publishes and what it keeps for itself. Regarding internal risks, the Eurobarometer of September 30 gives an order of magnitude, 15% of European employees report scams generated by AI and only 48% say they are capable of recognizing a fake video (brief). The editorial reading is summed up in a sentence. Europe has the most written framework, but the center of gravity of the month shifted towards incident management, which the AI Act addresses less directly than market placement. Two files remain open on ActuIA's radar. The bill number 3149 deposited by LFI deputies on September 15 was referred to the economic affairs committee, without being scheduled for the agenda so far (article). In Brussels, the KIDS Act begins its journey before the Parliament and the Council. Another deadline falls in November. The "Protective AI for Children" working group set up by high commissioner Sarah El Haïry must submit its proposals by the end of November, after two months of work (brief). ActuIA will also follow the review of its agents' actions that OpenAI says it is conducting since the Hugging Face incident, which will take months according to the company (brief). Same follow-up for the hearings of the Australian parliamentary committee, where Jason Kwon judged on October 6 that OpenAI's response was "not good enough," according to the Australian Financial Review (brief). The editorial will also monitor the possible publication of a founding act for the Super Intelligence Force. The next monthly report will be published early November.