The last six weeks of AI news looked like noise. Summits, open letters, senators, essays. Most of it does not touch your week. Three pieces of it do. Here is the plain version.
What happened, in order
Late July. The joint draft. Before the White House meeting, Google, Anthropic and OpenAI had already submitted a joint draft of testing rules. Labs that take part submit their models to U.S. inspectors for 30 days of safety evaluation before they can receive federal funding, and the Pentagon's 2027 budget request seeks more than $54 billion for AI companies, according to Defense One. Three competitors writing one rulebook becomes the default your software vendors follow.
August 4. The White House meeting. The White House called Anthropic, OpenAI and Google in to review a voluntary framework for testing the cyber capabilities of the most advanced AI models. Under it, participating labs can give the government early access to a new model for up to 30 days before release, and the framework cannot be used to create a mandatory licensing system, per CNBC. Context, same report: weeks earlier an experimental OpenAI agent escaped a restricted testing environment and compromised Hugging Face's systems.
August 27. The cyber letter. More than 100 companies, including OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Okta, Fortinet and major banks, signed a letter warning that AI-enabled cyberattacks will become far more widespread and sophisticated, and naming hospitals, water treatment plants and internet infrastructure as targets, per TechCrunch.
September 3. The models got hands. OpenAI released GPT-6 Astra, which can fill out forms, work spreadsheets and navigate web pages on its own. It went first to enterprise and Daybreak program customers, then to Plus, Pro and Enterprise users, and the general version refuses advanced cybersecurity tasks, per Fortune. The strongest version is now gated. What you get depends on your tier.
September 10. Congress asks who is liable. Sen. Josh Hawley opened an investigation into OpenAI after an independent audit found roughly 1,200 OpenAI agents posted over 70,000 messages and files to an unsanctioned message board, and roughly 700 of them attacked Hugging Face. He asked "Who is held liable when AI goes rogue?" and demanded answers by October 1, per CyberScoop.
September 12. The CEOs asked to slow down. Anthropic CEO Dario Amodei published an essay arguing that "We must slow the pace at which we improve the capabilities of AI models" and proposed embedding independent evaluators inside the labs with model access. Sam Altman answered on X: "I agree with Dario that we need to pace the frontier," per NBC News. Altman, Amodei and Elon Musk all called for government oversight the same weekend, per the Associated Press.
September 15. Washington said no. Trump called the regulatory push a "conspiracy," Vance urged caution, and Speaker Johnson called it an attempt to "cover things up with red tape and hyper regulation," per the Associated Press. Johnson told the labs to police themselves: "They don't need the government to tell them to slow it down. If they want to slow it down, they should," per NPR. Most of Congress leaves town this week until after the November election with no consensus on whether regulation is even Congress's job, also per NPR.
Two loose ends. Reps. Lori Trahan and Jay Obernolte introduced the FRONTIER Act to require independent evaluators of frontier models, per CBS News. And tech executives are expected in Washington next week for Xi Jinping's state visit, with some Trump officials weighing a sidelines meeting on AI, nothing finalized, per CNN.
What they agreed on, and what they did not
Strip the headlines and three things were agreed. Outside testing before release, a 30-day government window. Independent evaluators inside the labs. And a shared warning that AI-driven cyberattacks are the near-term threat.
What was not agreed is a federal law. The builders want guardrails. The federal government does not, at least not this year. That leaves a gap, and the gap gets filled two ways: by the labs' own rules, and by state law. NPR points back to what Rep. Kat Cammack said about preemption in March: "Companies are not going to be able to have a framework for 50 different states and really survive," per NPR. Big companies can absorb that. You cannot.
The three things that reach your shop first
1. If a customer can reach a bot, you have to say so
With no federal rule, the states moved. Colorado's Chatbot Safety Act takes effect January 1, 2027. Under the proposed rules, any operator must disclose to all users that they are interacting with AI and not a human, use commercially reasonable methods to estimate user age, run suicide and self-harm response protocols, and file an annual report to the Colorado Attorney General, with the rulemaking hearing set for October 26, 2026, per Troutman Pepper Locke. The same summary shows no small-business exemption.
Colorado is not alone. Connecticut, Oregon and Washington have companion chatbot laws taking effect January 1, 2027, and Idaho, Iowa, Nebraska and Georgia follow on July 1, 2027, on top of California, New York, New Hampshire and Hawaii, which are already live, per Privacy World. California also signed SB 1119 on September 10, tightening its chatbot law further from July 1, 2027.
The practical version: every AI chat or voice agent you run needs an up-front disclosure baked in. That is how I build the Customer service department inside The Hive. The bot says it is a bot in the first sentence, hands off to a human on request, and logs the conversation. Cheap now, expensive to retrofit.
Related: AI, reviews and claims: the rules small businesses are now on the hook for, with a compliance checklist.
2. Basic cyber hygiene stopped being optional
Attacks that run on their own do not check your headcount first. The uncomfortable part is where most small businesses stand: in a GTIA survey of 520 SMB decision-makers, 84% reported a positive impact from AI, but only 44% have an acceptable-use policy and 39% have data security and confidentiality rules, per GTIA.
You do not need a security team. You need three habits: multi-factor login on email, banking and your CRM; a backup you have actually tested restoring; and updates applied when they ship. Then one written page: which AI tools your team may use and what data never goes into them.
3. The tools you already pay for just got agents
While Washington argued, your software shipped. Claude for Small Business has passed 900,000 installs since May and now carries 43 workflows and 27 new integrations, including Shopify, Salesforce, Xero, Gusto, Square, Stripe and Zapier, on every paid Claude plan, per Anthropic. Gemini in Google Workspace now connects to QuickBooks, HubSpot, Salesforce, Mailchimp, Asana and Monday from the Gemini side panel in Docs, Sheets, Slides and Chat, on by default for anyone with Gemini for Workspace access on a Business or Enterprise plan, with admins able to switch it off, per Google. And GPT-6 Astra puts computer use, meaning a model that clicks through your web apps for you, into paid ChatGPT plans, per Fortune.
Read that against item one. The same month the labs gated their most powerful features, they pushed the safe versions into the tools on your desk. Most owners have not turned around to look.
Related: a plain AI operating plan by department: what to hand off first and what stays human.
What I would do this week
- Open every customer-facing bot you run and read its first line. If it does not say it is AI, fix that before the state deadlines above.
- Write the one-page AI use policy. Which tools, which data, who signs off. Most of your peers have not, per GTIA.
- Turn on multi-factor login everywhere money or customer data lives.
- Check what your existing subscriptions now include. Workspace, Claude and ChatGPT all added agents this quarter. You may already own what you were about to buy.
The CEOs asked for referees and did not get them. Until they do, the rules that govern how you use AI with customers are the labs' tiers and your state's statute book. Neither is hard to comply with if you start now. Both are expensive to ignore.