At the UN Security Council, Dario Amodei said his company would slow down as much as necessary to make AI safe. First person plural, no threshold, and no external check written into…

Meta's Muse promised an AI that phones businesses for you. People hung up on the synthetic voice, so external human operators made the calls while presenting themselves as the agent.

OpenAI asks for global technical standards for frontier AI and rules out licences, pre release review and approval. It cites aviation, where aircraft do not fly without certification.

An LNG carrier bound for Rovigo lost her control systems in the Adriatic and turned back. A suspected attack, not a demonstrated one, and an AI doing the boring half of targeting.

Amodei asks the industry to slow down, Altman and Musk agree within hours. No threshold, no verifier, no consequence. Non proliferation treaties have inspectors for a reason.

In the Red Sea the Italian Navy wins the engagements and loses the arithmetic. The obstacle is not the law of armed conflict. It is a procurement cycle that cannot buy software.

ChatGPT's Computer History records a working day into unencrypted Markdown files. The engineering is careful; the decision about who carries the consequences is not.

ShinyHunters claims the FBI personnel records, and says the route in was the HR platform. A claim, not an established fact, and a lesson about the surfaces nobody defends.

The Netherlands transposed NIS2 twenty two months late. Boards are now directly accountable, and Europe has one more national regime to reconcile.

When an AI agent causes damage, attribution runs through people and must be established one by one. In contract, the debtor must prove the failure was not attributable to it.