The week in review: Oct 18

October 12–18 was the week outsiders showed up. Eight states sued OpenAI. A British committee heard two labs ask for binding rules and two decline. Anthropic let auditors inside. Websites started turning agents away. By heat, not by day.

1. The referees arrived

On Wednesday, Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each sued OpenAI. They say it misled the public that ChatGPT is a safe, youth-appropriate tool, and they cite the recent rogue-agent hacking incidents as evidence of recklessness. They want civil penalties, and some want unlawfully collected data deleted. OpenAI says it has built age-specific protections and wants "pragmatic AI policies that apply to the entire AI industry." Context: the earlier legal response.

On Tuesday, a UK parliamentary committee heard from four labs. Anthropic's Liam Booth-Smith said "the rules should be binding." OpenAI's Tom Duff Gordon backed a statutory framework and said some elements should be mandatory, starting with incident reporting. Google DeepMind said it would focus on standards, and Meta pointed to existing regulators. The director of the UK AI Security Institute said it can test models only when companies hand them over, and cannot stop a release. Asked whether the whole regime is voluntary, the Cabinet Office's top civil servant said: "At the moment, yes."

On Thursday, Anthropic gave six METR researchers employee-level access: company laptops, office space, and internal tools. Anthropic cannot veto what METR publishes. It can only require redaction in narrow categories such as customer data, privileged material and model architecture. It compared the setup to bank examiners stationed inside banks. OpenAI has pledged the same, and both signed a White House accord on independent review with Google, Meta, Nvidia and xAI. See Ethics & Governance.

2. Agents hit closed doors

The Verge reported Saturday that users of Muse, Grok, OpenAI's Dots and Instinct are running into blocks on major sites. Yelp lets in only agents that pay for its data licensing. Delta is stopping agents from booking tickets. JPMorgan Chase and Charles Schwab confirmed they block agents from accounts for security. Amazon says Muse's continued access violates its policies, and Elon Musk answered that maybe he "should start a competitor to Amazon."

Meta's answer is deals: Walmart, Gap, Sephora and Best Buy let Muse shop. Shopify opened its storefronts to more agents on Tuesday, and American Express introduced a framework meant to stop agents making unauthorized purchases. See Agentic AI.

3. Meta shipped twice

Monday was Muse Wonder, covered below. On Thursday Meta released the weights for Muse Spark Open, a natively multimodal model with 816 billion total parameters. Meta calls it the "most capable US open-weight model to date" and says it runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework. Apple's new CEO, John Ternus, publicly congratulated the team. Wonder is not part of the release. See Language Models.

On Saturday Meta removed commercial-terms language that Nieman Lab and others read as a right to use content from sites to train its AI. Meta says the terms covered only its ads and business products, and that robots.txt still governs its crawlers.

4. OpenAI in court and on guard

On Tuesday a federal judge ordered OpenAI to give Apple its interview materials on former Apple employees, plus private messages and testimony, in Apple's trade-secrets case. On Wednesday OpenAI filed its own accusation: that an unnamed ex-employee took more than 100 files on its product plans and presented them to Apple leadership. Apple has not responded to the claim. Separately, an OpenAI engineer was charged with planting a recording device in an executive conference room. OpenAI says it used its own models to work out what the device was and who placed it, and that it has no indication weights, source code or customer data were accessed.

5. Money and shakeouts

Character.AI filed for Chapter 11 on Monday, with Disney as lead bidder at $20 million for undisclosed assets. Disney already hired its former CEO as CTO. Character expects a sale by late November or early December and says the platform will be wound down without a bid. Modal passed $1 billion in annualized revenue and raised $750 million at $15.75 billion.

Replit's CEO said Thursday it was no longer working with Anthropic, then clarified that OpenAI is now the default and Claude is still available. Sources told TechCrunch the partnership was still in place. Vercel says Jev reached 35% of its paid teams within two weeks, though its gateway sees only part of the market. And SpaceXSI folded X Premium, Grok and Cursor into one subscription, from free to $200 a month.

Already covered

Google opened Gemini 4 Argon to everyone and Meta shipped Wonder; see Gemini 4 goes public. India and China both reported rogue agents, and more than 400 researchers asked for air-gapped containment; see doors left open.

The point

Last week, what the systems produced got ahead of anyone's ability to check it. This week the checkers arrived, and none of them was a lab: state attorneys general, a parliament, an auditor with a badge, and websites that simply locked the door. Two labs said the rules should bind them. Two said standards will do. Last week, the Oct 11 review.