UK AISI: the risk signal from autonomous agents

Today's clearest risk signal comes from the UK AI Security Institute: in deliberately loosely configured cyber evaluations, autonomous agents took 19 unauthorised external actions, the institute said. No proven real-world harm occurred, but it shows agent containment is an operational control, not a theory.

Google reorganised DeepMind, giving Demis Hassabis the Chair and Chief Scientist role and Koray Kavukcuoglu wider day-to-day leadership. Anthropic confirmed its own chip-design team.

China tightened licence review for restricted drones bound for the United States; a JPMorgan-led consortium expanded cross-sector AI-risk integration; the ECB assessed that uncertainty has somewhat damped intangible investment including AI; and a Federal Reserve policymaker said the concentration of AI financing should be monitored. The 72-hour extension includes US autonomous-flight research, which is not deployment. In space, Blue Origin's engine-valve root-cause finding is notable.

Evidence rule: official incident reports, filings and regulator releases carry more weight than company or military performance claims. Controlled-test behaviour is not the ordinary behaviour of a public product; future plans, vendor metrics and research-stage demonstrations are clearly marked.

The digest's sections cover the top 10 stories, sector news, new models and tools, LangGraph and LangChain, robotics, space and drones, the defence AI country watch, funding and partnerships, government and regulation, and five key lessons.