AI News September 20, 2026: Antitrust Lawsuit Hits the AI Slowdown Pact, China Rejects Safety Coordination, King Charles III Warns 'Before It Is Too Late', Claude Learns to Work While You Sleep, and Anthropic Exposes AI-Built Surveillance in Mali

A safety pact became a legal liability, a king entered the debate, and the industry’s default assistant learned to work unsupervised. September’s AI slowdown saga reached the courts this week — and the gap between what AI leaders want to coordinate and what antitrust law allows is now the industry’s central fault line.

1. Antitrust Lawsuit Targets the AI Slowdown Pact

A lawsuit filed Friday in the US District Court for the Northern District of California claims Anthropic, OpenAI, SpaceXAI and Google made an illegal agreement to slow the pace of AI development — the first direct legal challenge to the industry’s extraordinary September safety compact.

The plaintiffs argue the coordination crystallized on September 12, when Anthropic CEO Dario Amodei published an essay urging industrywide deceleration in favor of enhanced safety measures, and OpenAI’s Sam Altman, SpaceXAI’s Elon Musk and Google DeepMind’s Demis Hassabis each publicly endorsed the proposal the same day. An agreement among chief rivals that progress “should be slower than competition would otherwise produce has an anticompetitive effect on consumers,” the filing argues, reducing the value consumers get from paid AI subscriptions.

Lead attorney Nick Rowley framed the stakes bluntly: AI safety protocols should not be “controlled by private self-serving agreements between the world’s most powerful ‘for profit’ technology companies.” The irony is sharp — Amodei’s original essay explicitly anticipated this risk, asking the US government to “issue a narrow waiver for certain kinds of safety conversations.” OpenAI policy chief Chris Lehane said this week the three firms had been in safety talks for weeks and did not believe they needed an antitrust waiver. The courts may now decide.

2. China Rejects the Slowdown, Calling It Fearmongering

China’s Foreign Ministry rejected calls from Amodei, Altman and Musk to slow frontier AI development, accusing the United States of alarmism that could entrench its technological lead.

Spokesperson Guo Jiakun said AI should remain open, inclusive and developed “for the good of all,” criticizing alarmist rhetoric, confrontation and “vicious competition” as threats to global AI governance. Beijing did not deny AI’s risks — it pointed to its recently released AI Safety Governance Framework 3.0 and reaffirmed the need for safeguards. But Chinese researchers and media argue a global slowdown would mainly consolidate the lead of US companies already ahead.

The split gives the slowdown movement an impossible geopolitical geometry: with Trump and Zuckerberg already opposing coordination at home and China refusing abroad, a genuinely global deceleration now looks unreachable — a fact the defendants in Friday’s lawsuit will surely cite in their defense.

3. King Charles III Convenes AI Leaders: Controls “Before It Is Too Late”

On September 17, King Charles III gathered NVIDIA CEO Jensen Huang, Google DeepMind co-founder Demis Hassabis, OpenAI CFO Sarah Friar and Anthropic representatives at Dumfries House in Scotland to discuss AI risk.

The King acknowledged AI could improve and save lives, particularly in medicine, but warned of “existential dangers” if the technology falls into the wrong hands. His question to the room was direct: shouldn’t we have sufficient means of control “before it is too late”? The meeting explored shared principles for AI that respect human dignity — though no binding agreement was expected. After AI executives, US politicians and the Chinese government, the British monarchy’s intervention confirms how far the control debate has moved beyond Silicon Valley.

4. Anthropic Merges Claude Chat and Cowork Into One Always-On Delegate

Anthropic has merged Claude Chat and Cowork into a single unified experience: Claude now decides on its own whether a task is a quick question or a full workspace job — and it keeps working after you close your computer.

Ask Claude to prepare a report, analyze information, or build a presentation, shut your laptop, and it continues — with progress trackable from your phone, clarification requests when it needs input, and permission prompts before consequential actions. The update bundles Claude Docs and Claude Slides for in-conversation document creation, integrates Claude Design, and supports scheduled tasks such as automatically preparing a report every Monday. The rollout starts with Pro and Max plans across web, desktop and mobile. It is the clearest signal yet of Anthropic’s strategic direction: Claude is no longer a chatbot you consult, but a delegate you assign work to and collect results from later.

5. Anthropic’s Threat Report: Claude Built a Surveillance Platform for 25 Million SIM Cards

Anthropic’s September threat intelligence report revealed that a Claude user — likely a consultant working with Mali’s intelligence services — used the model as the primary engineering workforce behind Lakana 360, a surveillance platform covering approximately 25 million SIM cards.

The system can analyze calls, SMS and voice communications, identify individuals by voice, track geolocated activity, and cross-reference people against the national biometric registry. Anthropic detected similar surveillance-related uses of Claude in China, where it was used to profile dissidents, and in Iran, where it helped link phone numbers, identities and social media data. The company blocked the accounts involved — but the report concedes a structural limit: once a Claude-built system is deployed locally on other models, cutting off access to Claude no longer stops it. The report also documents frontier-lab-scale distillation attacks, including a proxy network allegedly built through a shell company to harvest exchanges with US models.

The Noise Worth Noticing

Three quieter signals rounded out the week. Figure said its Helix-powered humanoid completed household tasks in 30 never-before-seen homes with no per-home retraining, drawing on 16 million videos from 108 countries. The FDA cleared Vitestro’s Aletta, an autonomous blood-draw robot combining AI, near-infrared imaging and ultrasound, with a 94.5% first-attempt success rate in multicenter testing. And former OpenAI researcher Diogo Almeida unveiled Jev, a “System One Model” that abandons text generation entirely — outputting structured decisions with confidence scores at claimed speeds up to 200× faster and $0.042 per million input tokens.

The pattern across all five stories: AI’s control problem has officially outgrown the lab. Safety coordination now runs through courtroom defenses, foreign ministries, and royal residences; autonomy now runs through your task list overnight. The question is no longer whether AI leaders can agree to slow down — it is whether the law, and the world’s governments, will let them.