On Monday in Tokyo, Japan's government opened a review of the AI law it passed in May last year. Kyodo and the Sankei Shimbun reported that the central question is whether an amendment should add penalties for malicious AI operators. The current law lets the state investigate and give guidance to AI businesses that harm people's rights, but it carries no penalties at all.

The review will map the risks behind it, including worse cyberattacks that use AI and systems that improve themselves fast enough to jump in capability. Digital Minister Toshiharu Furukawa asked members how Japan should make AI trustworthy. One attendee cautioned that "the world is not moving only toward tighter regulation." A working group led by university professors meets for the first time today, and the government will hear from developers before it sets out a direction by the end of the year. (Cabinet Office; Sankei; Kyodo via Tokyo Shimbun)

The briefing deck for the same meeting goes further than penalties. It starts from the premise that people carry final responsibility for AI, and asks how roles and authority should be redrawn as AI grows more capable and more autonomous. It asks whether every system needs a named responsible party, a decision trail that can be checked, a way for a person to step in and stop it, and routes to explanation, appeal and remedy.

The deck puts it in three questions: who does the task, who decides, and who answers for it. It also asks where responsibility should sit between people and the companies that build and supply AI, and what mechanisms would stop AI from breaking laws or running out of control. Those are questions a board can ask about its own systems today, without waiting for Tokyo. (Cabinet Office deck, Japanese)

So far, Washington has chosen guidance. NIST's super intelligence center, the National Security Agency and CISA published ten voluntary key practices on October 16 for handling incidents during the training and testing of advanced AI. They focus on cases where AI agents take significant actions nobody told them to take, actions that break the operator's controls or human oversight.

The practices run from preparation to reporting. Labs should set shutdown policies in advance, automatically stop the affected agent runs when a breach of controls is detected, and pause training if the incident is serious enough. They should also confirm containment works before restarting, notify the government, and share what they learned. It is voluntary, and it reads like a checklist a regulator could one day make mandatory. (NIST)

The US is also treating the AI supply chain as a security matter. In Tuesday's Federal Register, the Energy Department said it has added helium to its Critical Material List. It calls helium impossible to substitute in semiconductor manufacturing, leak testing of gas pipelines and quantum computing.

The department says demand is growing because helium enables advanced chipmaking, which it calls essential to super intelligence and data center supply chains. It ties the listing directly to the White House AI Action Plan. If your AI roadmap depends on chips, your suppliers' exposure to helium is now a question for your risk register, not just theirs. (Federal Register)

Already filed, if you want the run-up: The Owner (Korea's bank supervisor saying the bank owns every AI mistake), The Buffer (China telling employers to warn workers before AI takes their jobs) and The Lab (Britain's new AI biosecurity lab, funded partly in OpenAI and Anthropic credits).

Also on the shelf: The Clerk (a US appeals court taking a case away from a judge over an AI-drafted order), The Tab (Europe's Parliament fighting over who pays when AI does harm) and The Footer (Brazil's top court fining lawyers for hiding orders to its AI). Read today beside The Owner. Seoul offered its banks fewer rules in exchange for full responsibility. Tokyo is asking whether responsibility needs a penalty behind it.

Rules without penalties usually last until something goes badly wrong. In 1937 a US drug company sold Elixir Sulfanilamide, dissolved in diethylene glycol, a poison. No law required safety testing, and the company did none. The drug is believed to have killed 107 people.

The company's owner said he did not feel there was any responsibility on its part. The company paid a minimum fine under the 1906 food and drug law, and only for the label: an elixir had to contain alcohol, and this one did not. Congress passed the 1938 Federal Food, Drug, and Cosmetic Act, which required safety tests before a drug could be sold. The teeth came after the harm. (Elixir sulfanilamide)

Put these to management. If Japan, or any country we sell in, added penalties to its AI law tomorrow, which of our systems would be in scope, and who would answer for them? For each one, can we show who decided, and how a person can stop it?

Then the incidents. Do we have a written shutdown policy for AI agents, and has anyone tested it? Would we know within a day if an agent acted outside its instructions, and who would we tell, inside the company and outside it?

The Move

Japan's AI law has had no penalties. Its government is now asking whether that should change, and its own deck has already framed the questions of who acts, who decides and who answers. The US is writing playbooks for stopping agents and reporting incidents.

Before the next board meeting, ask the general counsel for one page: every AI system in a market where penalties are being discussed, its named owner, its shutdown path, and who it would report an incident to.

Build the answer before a regulator asks for it.

Seven bullets. One move.