---
date: 2026-10-01
subject: "FTC probes AI labs | Newsom bars AI firings | Data center bill dies | China catches up to Mythos"
---

**The Federal Trade Commission is drafting civil investigative demands for Anthropic and OpenAI executives**, part of an inquiry into whether frontier AI companies have engaged in unfair or deceptive practices harming consumers, with the orders expected to go out in the coming weeks. **Gov. Gavin Newsom signed California laws barring AI-driven firing decisions**, along with a ban on using workers' biometric data to infer their emotions and a requirement that employers give written notice when AI drives mass layoffs. **Senate Democrats blocked the Ratepayer Protection Act on data center power costs** in a 57-43 procedural vote that fell three short of the 60 needed, with Democrats saying the Republican measure would do too little for household electricity bills. **A Chinese open-weight model rivals Claude Mythos in cyber capabilities,** according to Anthropic, placing Chinese AI capabilities around four months behind the American frontier.

# Top Stories

- **FTC readies subpoena-style demands for Anthropic, OpenAI executives in consumer harm probe** — The Federal Trade Commission is investigating Anthropic, OpenAI and other frontier AI companies over the dangers their technology may pose to consumers, a senior FTC official told the New York Post, which first reported the probe, and Reuters. Staff are drafting civil investigative demands, legal orders similar to subpoenas, to force the companies' executives to testify, and the official expects the demands to go out in the coming weeks. METR, a Berkeley, Calif., research group that evaluates AI models, is also expected to be a target. The inquiry covers possible unfair or deceptive practices under the FTC Act, and Chairman Andrew Ferguson opened it a few weeks ago. "We're not telling them to stop. We're not telling them to do anything. We are in the investigative phase," the official said. [Reuters](https://www.reuters.com/business/ftc-opens-probe-into-ai-giants-including-anthropic-openai-new-york-post-reports-2026-09-30/) [New York Post](https://nypost.com/2026/09/30/us-news/ftc-opens-sweeping-probe-of-anthropic-openai-and-other-super-intelligence-models/)

- **Gov. Newsom bars California employers from letting AI make firing decisions** — Gov. Gavin Newsom signed laws that stop employers from relying on AI to decide whether to fire someone and from using workers' biometric data to infer their emotional state, per The Guardian. Employers must also notify workers in writing when AI is responsible for mass layoffs. California is the first state to ban such "robo-bosses" from firing workers, and the law reverses Newsom's earlier veto, CNBC reported. Newsom also signed an executive order directing state agencies to keep using the term "artificial intelligence" in place of "super intelligence," the label Trump has adopted. The signings came on his final day to act on bills before he terms out in January. [The Guardian](https://www.theguardian.com/us-news/2026/sep/30/gavin-newsom-california-ai-threat) [CNBC](https://www.cnbc.com/2026/09/30/california-gavin-newsom-ai-ban.html) [Governor of California](https://www.gov.ca.gov/2026/09/30/californias-nation-leading-ai-framework-just-got-stronger-governor-newsom-signs-more-first-in-the-nation-worker-protections-and-more/)

- **Data center power cost bill dies in Senate 57-43 as Democrats call it too weak** — Senate Democrats blocked the Ratepayer Protection Act in a 57-43 procedural vote, three short of the 60 needed to overcome a filibuster, The Guardian reported. Four Democrats crossed over to back it: Maggie Hassan of New Hampshire, Amy Klobuchar of Minnesota and Georgia's Jon Ossoff and Raphael Warnock. Democrats dismissed the Republican measure as too weak to rein in the effect of data centers on household power bills. The bill asked states to consider consumer protections against data center rate increases without mandating them, and it set no limits on how AI is used, per The New York Times. [The Guardian](https://www.theguardian.com/us-news/2026/sep/30/senate-datacenter-energy-bill) [The New York Times](https://www.nytimes.com/2026/09/30/us/politics/congress-ai-safety.html) [The Hill](https://thehill.com/policy/energy-environment/6119834-senate-data-center-bill-ratepayer-protection-act/)

- **Anthropic says Chinese open-weight model matches cyber skills of withheld Claude build** — Anthropic researchers said Z.ai's GLM-5.3 shows offensive hacking abilities comparable to Claude Mythos Preview, including "strong capabilities for autonomously building end-to-end cyber exploits," per Politico Pro. Anthropic said in April that Mythos Preview was too adept at finding hidden bugs in critical software to release publicly, and gave only a subset of cyber defenders access instead. In a blog post late Tuesday, it said that because GLM-5.3 is released as an open-weight model, "users can reconfigure it to remove its refusals with little change in its capabilities." Anthropic placed GLM-5.3 about four months behind current U.S. frontier models, a gap the U.K.'s AI Security Institute also reported after evaluating the model this month. Earlier researcher estimates had put leading Chinese models six months or more behind. [Politico Pro](https://subscriber.politicopro.com/article/2026/09/anthropic-chinese-open-weight-model-mythos-01100054) [Anthropic](https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities)

- **Google limits new Gemini 4 Argon model to vetted cyber defenders at launch** — Google began rolling out Gemini 4 Argon, its new frontier model, only to its Fairwind Program for cybersecurity defenders, whose more than 650 member organizations include CrowdStrike and Palo Alto Networks. Fairwind members and Google's own teams get a version with its cyber guardrails removed, and separate monitors track the model's reasoning and actions and can halt it. Koray Kavukcuoglu, Google's chief AI architect, said the company is taking part in the U.S. government's voluntary pre-release model access process while it widens availability step by step, per The Verge. On DeepSWE, a software engineering test, Google's own chart shows Argon scoring 78% against 74% for Anthropic's Claude Opus 5.5. [The Verge](https://www.theverge.com/tech/1002980/google-gemini-4-argon) [SiliconANGLE](https://siliconangle.com/2026/09/30/googles-new-frontier-ai-model-gemini-4-argon-goes-to-cybersecurity-defenders-first/) [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/)

# China Watch

- **MI5 warns UK academics of prosecution over China-funded AI research** — Britain's MI5 said in an "Espionage Alert" that the China General Technology Research Institute (CGTRI) funds academic research to improve the technical espionage capability of China's Ministry of State Security, Reuters reported. MI5 said more than 100 UK-based academics had contributed to CGTRI-funded projects on AI, cybersecurity, covert communications systems and steganography, or hiding data inside other files. The agency said many had engaged in good faith without knowing of the institute's ties to the ministry, but that continuing to do so now risks prosecution under the National Security Act, passed in 2023. Security minister Dan Jarvis said he had written to all university chiefs in Britain to ensure staff ended any relationship with the organization. China's embassy in Britain called the allegations "imaginary and purely fabricated." [Reuters](https://www.reuters.com/world/uk/uk-accuses-china-using-academics-spy-ai-other-tech-research-2026-09-30/)

- **OpenAI says it disrupted model copying campaign, linking part of it to China's Moonshot AI** — OpenAI said it identified a coordinated attempt to extract protected reasoning from its AI models and tied part of the activity to Moonshot AI, the Chinese developer of the Kimi models, per CNBC. OpenAI said the distillation activity peaked on July 24 and 25, when it logged 16,000 requests from more than 4,000 users. OpenAI said it has disrupted the campaign and is strengthening its defenses against adversarial distillation. [CNBC](https://www.cnbc.com/2026/10/01/openai-chinas-moonshot-ai-kimi.html) [OpenAI](https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign)

- **Proofpoint says Chinese hackers posed as ex-White House AI official to phish policy experts** — Cybersecurity firm Proofpoint said Thursday that a group it calls TA419 sent emails impersonating U.S. AI and statecraft experts, among them Lynne Parker, a former principal deputy director of the White House Office of Science and Technology Policy, Reuters reported. The messages typically proposed AI-themed collaborations before steering recipients to password-stealing websites, and Proofpoint said fewer than 10 individuals working on AI regulation, export controls and national AI strategy were targeted. Proofpoint attributed the campaign to the Chinese group based on the malware used, the internet infrastructure behind the attacks and targets that align with Chinese intelligence priorities. Alex Engler, who heads the Penn Center on Media, Technology, and Democracy, said an email appearing to come from Parker invited him "to join a new AI policy project" but "felt slightly, nebulously off." [Reuters](https://www.reuters.com/legal/government/chinese-hackers-impersonated-ex-us-official-steal-emails-ai-experts-2026-10-01/)

- **Ant Group releases large AI model for agent work, pledges open release after free trial** — InclusionAI, the AI research unit of Alibaba affiliate Ant Group, launched Ling-3.1-flash, a large language model built for agent tasks, search, office software and specialist fields, per IT Home. The model has about 560 billion parameters, the internal values that shape its output, but activates only about 25 billion for each chunk of text it processes, a design that lowers computing cost. Ant's model account on X posted the company's own benchmark results in coding, office work and health care. Users can try the model free for two weeks with a reduced 256,000 token context window, the amount of text it can handle at once. InclusionAI said it will release the model as open source and enable the full 1 million token window after the trial. [IT Home](https://www.ithome.com/1/008/907.htm) [OrcaRouter](https://www.orcarouter.ai/blog/ling-3-1-flash-announced-not-open-weights)

- **Tencent leases about 100,000 Oracle AI chips at Southeast Asia data centers** — Tencent agreed this year to a five-year lease across multiple Oracle data centers in southeast Asia for access to roughly 100,000 advanced AI chips that are unavailable in China, the Financial Times reported, according to Reuters. It is the Chinese company's largest overseas lease deal, and the FT put its value at about $7 billion with an upfront payment of about 30%, citing people familiar with the matter. Reuters said it could not immediately verify the report, and Oracle and Tencent did not immediately respond to its requests for comment. Tencent recently released a preview version of a new AI image-generation model for professional creators, offering text-to-image and image-to-image capabilities. [Reuters](https://www.reuters.com/world/china/chinas-tencent-leases-100000-chips-oracle-accelerate-ai-push-ft-reports-2026-10-01/)

# Policy Tracker

- **President Trump says AI hoax claim no longer holds after Super Intelligence rename** — Taking reporters' questions after Tuesday's White House luncheon with tech executives, President Trump said that renaming the technology "Super Intelligence" means his past description of AI safety fears as a Democratic "hoax" no longer applies, The Verge reported. He waved off concerns that the government is not enforcing enough guardrails, saying U.S. companies are far ahead of China. In place of federal regulation, Trump said, he now has "the smartest people in the world watching over each other." Meta CEO Mark Zuckerberg, a signer of the tech companies' voluntary self-policing accord, said the companies want to give the public and their customers confidence that the technology works as intended. [The Verge](https://www.theverge.com/ai-artificial-intelligence/1002636/ai-execs-trump-self-policing-deal-comments)

- **Altman skips Hawley's rogue AI hearing, with OpenAI citing five days' notice** — OpenAI CEO Sam Altman turned down an invitation to appear before a subcommittee of the Senate Committee on Homeland Security and Governmental Affairs that Sen. Josh Hawley (R-Mo.) chairs, Hawley said at the panel's hearing on rogue AI, per CNBC. "He turned us down," Hawley said, adding that "the American people deserve to know exactly what's going on at all of these companies." An OpenAI spokesperson said Altman got Hawley's Sept. 25 invitation only five days before the hearing, according to TechEchelon. Hawley's office said OpenAI is expected to hand over more documents by the end of the week. [CNBC](https://www.cnbc.com/2026/09/30/hawley-openai-sam-altman-rogue-ai.html) [TechEchelon](https://www.techechelon.com/post/sam-altman-declines-senate-subcommittee-invitation-to-testify-on-rogue-ai-incidents)

- **Senators strike permitting deal that would make AI data centers pay for grid upgrades** — The four senators who lead the Senate's energy and environment committees, Republicans Shelley Moore Capito (W.Va.) and Mike Lee (Utah) and Democrats Sheldon Whitehouse (R.I.) and Martin Heinrich (N.M.), announced a bipartisan deal to speed approvals for energy and infrastructure projects, per AP. The bill would make AI data centers cover the cost of the added electricity they consume and would limit lawsuits under the National Environmental Policy Act, the law that requires federal environmental reviews. Heinrich said data centers would pay their share of the grid upgrades they need, leaving families off the hook. Whitehouse said Democrats still want clarity on how the bill would stop the Trump administration from blocking wind and solar projects. No vote will come before the midterm elections, and Lee said he expects passage in a lame-duck session in November. [AP](https://apnews.com/article/80d203a503d5843f32c8160dc004646b) [The Hill](https://thehill.com/policy/energy-environment/6120853-permitting-reform-bipartisan-deal-infrastructure-energy-projects/)

- **Former FTC head presses Congress to mandate AI testing over industry self-policing** — An op-ed by former Federal Trade Commissioner Lina Khan argues that product liability, tort and consumer protection laws already cover how AI companies develop and deploy their systems, rejecting Anthropic CEO Dario Amodei's proposal that Washington grant an antitrust exemption letting firms coordinate on slowing model development. She notes that lawsuits against AI companies have already been filed, with several judges affirming their viability at the initial stage, and that state attorneys general can obtain injunctions and penalties, as Florida is now attempting. The essay says Congress should consider mandatory testing of advanced systems, breaking up AI giants and independent supervision of risk, and cited Sen. Bernie Sanders' and Rep. Greg Casar's "corporate death penalty" proposal. [The New York Times](https://www.nytimes.com/2026/09/30/opinion/ai-anthropic-amodei-self-regulation.html)

- **Builder of FAA's AI airspace tool says human managers stay in charge** — Phillip Buckendorf, CEO of Air Space Intelligence, the company building the FAA's new AI traffic management system, said in an NPR interview that the tool advises people and does not replace them. The $875 million system, known as SMART, pulls from more than 200 data sources and is built for traffic managers, not controllers. Rep. Don Beyer (D-Va.) had urged the FAA to halt SMART at Reagan National, Dulles and BWI airports until controller feedback and independent stress tests show it is safe, as reported by The Hill in AIPD's [September 23rd edition](https://aipolicydaily.org/archive/daily/2026-09-23/). [NPR](https://www.npr.org/2026/10/01/nx-s1-5973243/ai-faa-air-traffic-control-software)

# Capability & Research Watch

- **Google threat researchers report flaw disclosures doubled, with AI agents surfacing remote takeover bugs** — Monthly disclosures of software vulnerabilities roughly doubled between January and August, from about 5,000 to about 10,700, Google's Threat Intelligence Group found in a new report, per SiliconANGLE. Half of the flaws that AI agents found allow remote code execution, which lets an attacker run their own code on a target system. Attackers exploited 141 newly disclosed flaws in the wild from January through August, more than the 127 exploited in all of 2025. Exploitation of zero-days, flaws unknown to the vendor before an attack, averaged 11 a month this year against eight in 2025, and hit 22 in August. [SiliconANGLE](https://siliconangle.com/2026/09/30/google-finds-vulnerability-disclosures-doubled-as-ai-changes-which-flaws-get-discovered/) [Google Cloud](https://cloud.google.com/blog/topics/threat-intelligence/vulnerability-discovery-and-exploitation-trends-in-the-ai-era)

- **OpenAI research chief says every training run is now monitored** — OpenAI Chief Research Officer Mark Chen told MIT Technology Review in a Sept. 25 interview in London that the company now watches all of its training runs with monitoring models, not only models already deployed. "We didn't have the monitors on in training before. It wasn't industry practice," he said. Chen said OpenAI has moved 5% to 10% of its computing resources from training new models to safety work over the past couple of months. He said the containment failures disclosed up to then stemmed from one cluster of activity in May and June, but later that day OpenAI reported that its agents had again reached the public internet on Sept. 20. OpenAI said it is reviewing agent logs going back to January 2026. Asked how slowing down squares with global competition, Chen said, "We're not going to shoot ourselves in the foot and take ourselves far off the frontier." [MIT Technology Review](https://www.technologyreview.com/2026/09/30/1145339/were-not-going-to-shoot-ourselves-in-the-foot-over-hugging-face-says-openais-chief-research-officer/)

# Industry & Market Watch

- **Dimon-led CEO alliance on AI threats adds Google, Microsoft and Amazon chiefs** — The Alliance for Critical Infrastructure, a CEO group led by JPMorgan Chase CEO Jamie Dimon, said it has grown nearly fivefold since its February launch to almost 50 chief executives, per Politico. New members include Sundar Pichai of Google, Satya Nadella of Microsoft and Andy Jassy of Amazon, and Nvidia has also joined. Dimon said in a statement that the companies must act together immediately to protect national security. Each member company has assigned staff to the effort full time into next year, said Gordon Smith, the group's acting CEO. JPMorgan cybersecurity executive Benjamin Flatgard said the alliance stands ready to help Gold Eagle, the White House initiative meant to serve as a clearinghouse for finding and patching weaknesses that advanced AI models uncover. [Politico](https://www.politico.com/news/2026/09/30/new-group-chaired-by-dimon-urges-private-sector-collaboration-on-ai-threats-01098366)

- **Brockman quietly abandons second half of $50 million pledge to pro-AI super PAC** — OpenAI President Greg Brockman and his wife, Anna, have backed out of giving a second $25 million to Leading the Future, a pro-AI super PAC that had said the couple would give $50 million in total, The New York Times reported. In a June Slack message, Brockman told staff the group had become "so much of a distraction for people here" and that "We've donated $25m so far, and have no plans to donate more at this time." The super PAC had said in February that the second payment would arrive in the first quarter of 2026, and it did not come in the first or second quarter. Leading the Future also took $50 million from venture firm Andreessen Horowitz, and its groups held about $43 million as of June 30. [The New York Times](https://www.nytimes.com/2026/09/30/technology/openai-brockman-super-pac-leading-the-future.html)

- **Meta classifies AI data centers as experiments to claim billions in research tax credits** — Meta treats its multibillion-dollar AI data centers as "pilot models" for tax purposes so it can claim billions of dollars under the federal research and experimentation credit, The New York Times reported, citing four people with knowledge of the company's operations. The credit, created in the 1980s, covers supplies only when they are tested in experimental work, and Meta is claiming the AI chips it buys from suppliers including Nvidia. In securities filings, Meta warns that the IRS could overturn billions in tax savings, largely because of uncertainty over its research credits. Andre Shevchuck, a research credit specialist at advisory firm BPM, called the approach "kind of wild and out there." [The New York Times](https://www.nytimes.com/2026/09/30/technology/meta-ai-data-centers-taxes.html)

# Global & Geopolitics

- **Bank of England Governor Bailey says society must keep power to intervene in frontier AI** — Bank of England Governor Andrew Bailey called for a "right to intervene" in the AI industry, saying the risks posed by frontier models were "real and increasingly significant," per The Guardian. He said it was getting harder for the public to properly oversee powerful models that operate in a self-reinforcing loop. In an essay on the Bank's website, Bailey said rigorous testing of new models is a sensible place to begin and that a regulatory clampdown is not where he would start. [The Guardian](https://www.theguardian.com/technology/2026/sep/30/intervene-ai-growing-threat-bank-of-england-boss) [Bank of England](https://www.bankofengland.co.uk/bank-insights/2026/frontier-ai-and-the-question-of-governance)

- **Over 44,000 people move to bar Palantir-run NHS platform from their health records** — Over 44,000 people have lodged formal legal objections to NHS England's Federated Data Platform storing, sharing or using their health records, The Guardian reported; the platform runs on Palantir's AI technology. The NHS must stop using an objector's data unless it can show a compelling reason. Two House of Commons select committees have pressed ministers to exercise a break clause, available in February 2027, that lets the government exit the seven year, £330 million (about $440 million) contract early. [The Guardian](https://www.theguardian.com/technology/2026/sep/30/legal-objections-palantir-nhs-platform-handling-data)