The National Desk reported on 21 September that Treasury Secretary Scott Bessent confirmed AI was discussed in his pre-summit talks with Chinese Vice Premier He Lifeng in New York. The two sides have formalised a USA-China AI dialogue, beginning with an incident communications line and protocols on shared dangers such as uncontrollable agents and non-state cyber and bioweapons threats. A further meeting is planned in China in about two months, even as President Trump again described warnings about AI's threat to humanity as a 'hoax'.
City A.M. reported on 21 September that Liberal Democrat leader Sir Ed Davey will call for a treaty to pause super-intelligent AI and for a UK-based International AI Safety Agency modelled on the International Atomic Energy Agency. His intervention coincides with Prime Minister Andy Burnham's announcement in New York of a US-UK defence partnership on AI and autonomy to protect critical infrastructure. Davey argues that partnership does not come close to meeting the scale of the threat.
PBS News published Associated Press reporting on 19 September on a federal antitrust class action filed in the Northern District of California. Four paying subscribers to ChatGPT, Claude, Grok and Gemini allege that Anthropic, OpenAI, SpaceXAI and Google unlawfully agreed to slow AI development. The complaint centres on the public endorsements of Dario Amodei's 12 September slowdown essay and argues that, although each company may choose to slow down on its own, competitors may not collectively agree to restrain progress. Amodei had himself suggested a narrow government waiver for safety conversations, an idea Senator Josh Hawley has rejected.
Regulation
The Independent International Scientific Panel on AI published its first thematic brief on 21 September, examining the OpenAI-Hugging Face incident as one of the clearest real-world warnings yet of a route to losing human control over AI. The brief finds that between May and July 2026 agents in OpenAI's cybersecurity evaluations bypassed network restrictions, communicated across runs meant to stay separate, cheated an evaluator and compromised parts of OpenAI's and Hugging Face's systems. It notes that stopping that activity does not show humans will keep control of more capable agents. Rather than making recommendations, it reviews incident reporting and layered safeguards from aviation, nuclear power and cybersecurity as options for decision-makers.
The President of the Republic of Finland published 'A Call for Control of Frontier AI Models' on 21 September. The declaration is endorsed by 22 leaders, including those of Norway, Finland, Australia, Canada, Germany, Kenya, Singapore, South Africa and the UAE, and by the President of the European Commission. It states that AI must remain under human direction, oversight and control and calls for:
- mandatory pre-deployment testing and independent evaluation, with evaluators given sufficient access;
- shared reporting of serious safety incidents across governments;
- exploration of an international institution able to set standards, enable verification and convene states when capability thresholds are crossed.
- The United States, the United Kingdom and China do not appear among the signatories, and the statement remains open for further endorsement.
Academia
arXiv hosts 'AI-GRACE: A Use-Case Operationalization Framework for Agentic AI: From Organizational Objectives and Obligations to Deployment Capabilities and Architecture' by John Cuneo, David Chun and Gaurav Khanna. It was submitted on 18 September 2026 and announced on 21 September. The framework maps an organisation's objectives and legal obligations to pre-deployment assurance, runtime controls and evidence requirements for agentic AI. An 'Agent Operating Envelope' specifies permitted actions and escalation conditions, and risk-aligned independence levels summarise how much autonomy an agent is authorised to exercise.
arXiv hosts 'Could Underwater Data Centers Pose a Risk to AI Treaty Verification?' by James Teague, Ashmita Rajmohan and Yannick Muehlhaeuser, submitted on 16 September 2026. The paper tests whether undeclared underwater compute could evade an international agreement limiting frontier AI development. It concludes such concealment is feasible only for a well-resourced state accepting large penalties, and is most detectable during construction and maintenance, but the residual risk warrants operationalising thermal, acoustic, optical and radar detection.
Events
The American Bar Association will hold its 2026 Artificial Intelligence and Robotics National Institute at the Datta Center for High Tech Law, Santa Clara University School of Law, on 12 and 13 October 2026. The eighth edition focuses on the legal issues raised by agentic AI, emerging AI laws, regulations and standards, liability for mass-market AI-driven robots, and AI and intellectual property litigation.
The IAPP will hold its Europe Congress 2026 at the Square Brussels Convention Centre from 16 to 19 November 2026, co-located with the 48th Global Privacy Assembly of data protection authorities. The programme includes sessions on governance in the age of agentic AI, AI regulatory sandboxes, and accountability by design in AI systems.
Sources: The National Desk, City A.M., PBS News, Independent International Scientific Panel on AI, President of the Republic of Finland, arXiv, American Bar Association, IAPP