AI News October 6, 2026: OpenAI Puts Visual Ads Inside ChatGPT, AI Giants Testify Under Oath in New York, Claude Cowork Goes Cloud-Only

The AI industry’s two biggest pressures — making money and facing accountability — collided within 24 hours. OpenAI is now openly building an ads business inside humanity’s most-used AI product, while four frontier labs were forced to raise their right hands before a city council. Meanwhile, Anthropic quietly made agents cloud-mandatory, OpenAI named names in an industrial-scale model theft attempt, and Meta’s models are doing mathematics research. Here are the five stories that matter today.

1. OpenAI Brings Visual Ads to ChatGPT Image Generation

On October 5, OpenAI announced a new visual ad format for ChatGPT that appears during image generation, with US testing beginning later this month among an initial group of advertisers. The company says the ads are clearly labeled, sit separately from the generated image, and — critically for trust — advertising does not influence the image result itself.

The bigger story is the measurement infrastructure shipping alongside it. OpenAI is connecting Hightouch, Tealium, and LiveRamp so advertisers can pipe first-party conversion signals directly into ChatGPT Ads, and its attribution partner roster now spans AppsFlyer, Triple Whale, Adjust, Rockerbox, Northbeam, Branch, Singular, Kochava, and others. Placement guardrails assess whether a conversation is appropriate for advertising, keeping ads out of “emotionally vulnerable” contexts.

The context: ChatGPT now reaches 1.2 billion people weekly, ads are live in 38 countries, and Amazon struck a deal last week letting brands on its ad platform buy ChatGPT placements. With a reported $1B+ ad run rate and a 2026 IPO in motion, OpenAI is behaving less like a research lab and more like the next great media platform.

2. Historic First: AI Executives Testify Under Oath in New York

At 11 a.m. Monday, the New York City Council’s rare Committee of the Whole — all 51 members — convened the first sworn public testimony from the major AI labs on AI risks. Anthropic sent Logan Graham, head of its Frontier Red Team; OpenAI sent Morgan Dwyer, head of policy development and operations; Google sent Alice Friend, director of AI and emerging tech policy; Meta sent AI policy director Shane Cahill.

Elon Musk’s SpaceXAI was subpoenaed but declined to appear; the city is now pursuing legal action. The hearing examined Speaker Julie Menin’s 10-bill package, including mandatory kill switches, third-party validation before deployment, a first-in-the-nation AI whistleblower incentive program, and a private right of action for New Yorkers harmed by AI agents. Former Anthropic researcher and whistleblower Jacob Coxon also testified. With the federal government endorsing self-regulation, Menin’s message was blunt: “New York City needs to take matters into its own hands.”

3. Claude Cowork Goes Cloud-Only — Starting Today

A quiet but consequential change hit Anthropic’s Pro and Max users on October 6: new Cowork tasks now run in Anthropic’s cloud by default, and the “Only on your computer” option has been removed from settings. Tasks already running locally stay local, and tasks touching files on your machine still require the desktop app open — but if you want guaranteed local execution, Anthropic now points you to Claude Code.

The trade-off is classic: cloud execution means 24/7 scheduled agents that never sleep and never depend on your laptop being awake, at the cost of your files being processed on Anthropic’s servers. For privacy-sensitive enterprise workflows, this is the day local-first agent execution on Claude became opt-in elsewhere.

4. OpenAI Exposes “Adversarial Distillation” Campaign Tied to Moonshot AI

In a September 30 security disclosure, OpenAI revealed it disrupted a coordinated campaign to extract protected reasoning from its models — what it calls “adversarial distillation.” The campaign began quietly on July 1, spiked on July 24–25 with 16,000 extraction-pattern requests from more than 4,000 accounts, and by July 28 OpenAI had disrupted a related cluster of more than 15,000 accounts. The company notes these figures describe attempted, not necessarily successful, extractions.

OpenAI attributed a core cluster of the activity to individuals associated with Moonshot AI, the Chinese lab behind the Kimi models, while cautioning it is unclear whether all operators belonged to one actor. The disclosure — coming three weeks after researchers published the “Stealing Reasoning Traces from Proprietary LLM APIs” attack — signals that frontier labs now treat model distillation as a security incident, not a gray-zone research practice.

5. Meta’s Muse Spark Co-Authors Papers on Open Math Problems

Per AI Weekly’s roundup, Meta AI released six mathematics papers co-authored with human mathematicians using Muse Spark 1.1 and 1.2 in Thinking Mode through the standard Meta.ai chat interface — and five of them address previously open problems, including a sharp threshold for high-dimensional ellipsoid fitting, finite-time wave collapse in biharmonic nonlinear Schrödinger equations, and a counterexample disproving the García-Martínez–Pérez-Rodríguez conjecture. Peer review is pending, but the result would be a milestone for AI-assisted discovery — and a fitting coda to a month in which Meta shipped Muse Spark 1.3 and launched the Muse personal agent.

The through-line this Tuesday: the industry is simultaneously commercializing (ads), localizing accountability (sworn testimony), centralizing its agents (cloud-only Cowork), and fortifying its moats (distillation enforcement). The race to monetize AI is now officially as aggressive as the race to build it.