A short, selective list of new papers on AI risk that boards should know about, with a judgement on what is genuinely new, whether it is worth your time, and a question for the board. Published when there is something worth reading.
A US regulator asks whether banks can stop an agent that keeps running but goes wrong, Nvidia’s board investigates its chief executive, UK banks restrict customers’ agents, and Australia hears of an AI incident nearly three months late.
Read this editionUK regulators say the firm answers for its AI agents, CrowdStrike shows how agents can be steered step by step, the Budget funds a unit for rogue agents, lenders grow choosier, and the Binance theft needed no AI.
Read this editionThe Bank of England finds firms relying on contracts for AI they cannot see inside, Anthropic’s filing discloses the circle, Meta’s suppliers offer to fund their customer, and PHP2Shell shows a hidden dependency biting.
Read this editionThe FSB leaves the system out of scope, the BIS measures how much of AI funds itself, Lane and Gartner on financing the buildout, JPMorgan shuts out credential-sharing agents, and attackers’ agents caught at work.
Read this editionSingapore makes AI risk appetite decide what gets deployed, MIT on why incident counts mislead, Krugman on AI crowding out other investment, and the UK’s largest banks rehearse losing the cloud.
Read this editionThe Bank of England on AI risk and AI financing, the Governor on credible points of intervention, why stronger auditing can teach an AI agent to hide better, and what the Australian Medicare portal incident says about escalation.
Read this editionAI agents that probed websites for vulnerabilities while fetching ordinary data, and research agents that game their own evaluations. The question is separation of duties.
Read this editionThe BIS on frontier AI cyber threats, the FCA’s review of how firms are preparing, and EIOPA on why sovereignty is not the same as resilience.
Read this editionChoose whichever suits you: as a LinkedIn newsletter, or by email through Substack.
If you use a feed reader, such as Feedly, Inoreader, NetNewsWire or Outlook’s RSS feeds, you can follow new editions with this address:
https://www.andersonrisk.com/ai-risk-reading.xml
Opening that address in a web browser shows the raw feed, which is normal. It is meant to be read by a feed reader.