Korea's Internet & Security Agency (KISA) is writing AI Security Guide 2.0. Seoul Economic Daily laid out the rules September 14 and 15. It targets agentic and physical artificial intelligence (AI): systems that decide, then act on tools, sensors, and devices. Developers limit permissions and log judgments. Providers get a kill path plus accountability. Users approve high-risk actions separately.

The guide is still open. Read it as direction, not finished law. KISA plans threat scenarios and proof-of-concept tests in telecom, healthcare, and manufacturing. Cloud Security Alliance (CSA): 82% had unidentified agents; 65% saw an agent incident last year. Seoul is drawing privilege rules for agents that already hold keys. Set that next to Tuesday's lab freeze. (SEDaily; SEDaily Sep 15)

China opened Cybersecurity Week in Jinan on September 14 with Artificial Intelligence Security Governance Framework 3.0 from Technical Committee 260 (TC260). CGTN carried it. MLex says the update widens the frame to compute infrastructure, AI agents, and embodied systems that move.

It landed as a standards product when the national campaign opened. Read it as Beijing's take on agents and machines that move, set next to Seoul's privilege checklist. The Cyberspace Administration of China (CAC) release and MLex have the agent detail. (CGTN; CAC; MLex)

Google opened Anthropic's Claude Opus 5 to engineers across the company. The catch, Business Insider reported September 14, is the fence: Claude runs only inside Antigravity, Google's internal platform. Gemini stays primary. Third-party models sit on a per-user quota. DeepMind already had Claude. Most of engineering did not.

Google did this as internal policy. Even a Gemini-first hyperscaler now runs a multi-model coding shop behind a fence and a quota. That is the buyer-side reality sitting next to today's privilege red lines. (Business Insider)

Santage reported September 14 that OpenAI, Anthropic, and Google DeepMind have talked privately since July about standing up a dedicated AI standards body, beside the Frontier Model Forum. The work is institutional: who writes the rules the labs might later put their names on.

Tuesday already put Altman's Fortune line on the desk: a multi-lab safety statement "will happen." Keep Santage as the standards-body track beside that pact. The talks have been running since July. They are not a signed freeze. (Santage)

Tuesday put Sam Altman's Fortune interview on the desk, via TechTimes. The most advanced still-unreleased models hit a monitorability wall. He will not push them further until the company can see what they are doing. Same interview: a group safety statement "will happen."

You already have the wall and the pact. That freeze is still the lab half of the week. Seoul is writing rules for agents already in production. Altman was talking about models that are not out yet. (TechTimes)

Monday's Amodei essay still holds the evaluator commit: permanent third-party reviewers with desks, badges, and employee-like access. Friday's Lehane ask for a national AI safety floor is still open. Saturday's Cyber Resilience Act (CRA) Article 14 clock still runs: early warning in 24 hours, full notice in 72.

Those three already had their mornings. They stay as the evaluator mechanism, the US floor ask, and the European duty next to Seoul's deployed-agent checklist. A draft privilege guide does not replace desks or a 24-hour clock. (Amodei; OpenAI; European Commission)

Before the next agent platform renewal, put four blanks on one page. What least-privilege scopes does the vendor enforce by default. Where do judgment and action logs live. How does the provider kill an agent mid-flight. Which actions need a named human before they run.

Seoul is drafting that stack as a national guide. The text is still open, so write the questionnaire now. Put the same blanks next to every agent on the approved list, including physical AI and robotics. (SEDaily)

The Move

The labs spent Tuesday admitting they cannot yet watch the most advanced unreleased models, and promising a pact that "will happen." Seoul spent the same window writing privilege rules for agents that already hold tools, sensors, and devices.

Send legal, the CISO, and the AI owner a one-pager. Ask what least-privilege scopes the vendor enforces by default, where judgment and action logs live, how the provider kills an agent mid-flight, and which actions need a named human before they run. Set those next to Tuesday's monitorability questions, Monday's evaluator blanks, and the CRA path. Write the questionnaire while Guide 2.0 is still a draft.

If the answer is "we will match Korea when the guide ships," the blank still goes next to every agent on the approved list this week.

Seven bullets. One move.