Daily AI brief · September 19, 2026
Trump and Newsom Split Hard on AI Safety
California ordered work on a frontier-model kill switch a day before the President called AI safety a hoax, while Anthropic and Accenture pledged $1 billion each for embedded model evaluators.
Trump promises an "AI Force" and an AI czar, calls safety fears a hoax (Axios, NBC News, TechCrunch) In a post on Friday, the President said he will create an AI Force modeled on the Space Force and appoint an AI czar, and called AI alignment concerns a "hoax." He wrote that the administration will not "hinder or stifle" the industry. No appointee was named and no details were given on what the AI Force would do. What it means for SMBs / MSPs: Don't expect federal rules to set your clients' AI guardrails any time soon; your own policies will have to do that work.
Newsom orders California to draft kill-switch and audit rules (CNN, Washington Post, Bloomberg) Governor Newsom signed an executive order on September 18 directing a working group to deliver recommendations within two months. Options include requiring a kill switch for frontier models and having independent third parties draft AI companies' safety plans. In 2024 he vetoed a bill that would have required a shutdown capability. What it means for SMBs / MSPs: Rules like these land on model makers first, but the vendors your clients buy from will pass the paperwork down in contracts.
Anthropic and Accenture pledge $1 billion each for embedded evaluators (Anthropic, TechXplore) Both companies expect to spend at least $1 billion over five years on independent evaluation of frontier AI. Faculty, Accenture's AI business, will lead red-teaming, alignment assessments and safeguard testing, with evaluators given access inside Anthropic comparable to an employee's. The deal is non-exclusive, and Anthropic says it is also talking with METR and other nonprofits. What it means for SMBs / MSPs: Third-party testing is becoming something a vendor can point to, so start asking your AI vendors who tests their models and whether you can see the results.
Claude Code reads AGENTS.md when there's no CLAUDE.md (DevOps.com, DEV Community) Claude Code 2.1.277 falls back to an AGENTS.md instruction file in projects that have no CLAUDE.md in the folder or any parent. It's a fallback, not a merge: if CLAUDE.md exists, AGENTS.md is still ignored. It isn't available yet through Bedrock, Vertex or Foundry. What it means for SMBs / MSPs: Teams using more than one coding agent can keep one set of project instructions instead of two that drift apart.
ChatGPT's desktop browser now runs Chrome extensions (Runtime Wire) OpenAI added Chrome extension support to the browser built into the ChatGPT desktop app. Admins can set site rules for access, uploads, downloads and browser control that users can't override, and work workspaces may need an admin to enable it. What it means for SMBs / MSPs: That's a second browser on your clients' machines, with its own extensions and passwords, so bring it under the same policy as Chrome.
One instruction file for every coding agent. This is for champions and MSPs whose teams use Claude Code alongside Codex or another agent. It takes about 20 minutes. First step: in one repo, move the shared project rules into AGENTS.md, delete or rename the CLAUDE.md, update Claude Code, and ask each agent to summarize the project rules back to you. If both answers match, roll it out.
Write a one-page AI use policy this week instead of waiting for federal rules. The President called AI safety fears a "hoax," while California's governor ordered a two-month study of kill-switch and audit rules, so no single rulebook is coming soon. Cover approved tools, what data can go in, who approves new agents, and what to ask vendors about who tests their models.
*For MSPs:* A two-page AI use policy gives owners who were waiting on regulation a clear place to start.
Reading "kill switch" as law. Newsom's order starts a two-month study, not a requirement. Don't rewrite contracts or pause projects over it; watch for the working group's recommendations.
AI oversight is being privatized. With federal policy moving away from safety rules, the checks are coming from states, from lab-funded evaluators, and from the labs' own disclosures. For small businesses, that means vendor due diligence matters more: the evidence you need about a model will increasingly sit in a vendor's evaluation reports, not in a regulation.
Compiled on September 27, 2026 from reporting published on or before September 19, 2026.