Anthropic seeks rehearing | AI firms split on UK law | Senators urge AI tests

Bottom Line Up Front

Anthropic asked a federal appeals court to revisit its Pentagon blacklisting, challenging a 2-1 September ruling that upheld a national security label barring the company from defense work. OpenAI and Anthropic executives urged British lawmakers to pass binding AI rules at a House of Commons hearing, while Google DeepMind and Meta declined to back a statutory framework. Sens. Maria Cantwell and Mike Rounds pressed the administration to expand federal AI model testing and speed up work on safety standards. Bank of England and International Monetary Fund chiefs warned that AI agents and fraud erode trust in money, saying rising scam costs could drive people out of the formal financial system.

I.Top Stories

Anthropic asks D.C. Circuit to rehear Pentagon supply chain risk case

Anthropic asked the U.S. Court of Appeals for the D.C. Circuit to reconsider a September panel decision upholding the Pentagon's national security supply chain risk label for the company, which blacklists it from defense business, per Reuters. The panel ruled 2-1. In its petition, Anthropic argued the majority turned a federal law "aimed at hostile foreign actors into one that presumptively reaches every government contractor." The company also said the court let the Pentagon act without any pre-designation process despite evidence of retaliation for its speech on AI safety. Anthropic has said the designation has cost it billions of dollars in lost business.

Read at Reuters ↗ • Read at Bloomberg ↗

OpenAI and Anthropic back binding UK AI rules as Google DeepMind and Meta hold back

Executives from OpenAI and Anthropic told a House of Commons committee that voluntary pledges no longer suffice and that Britain should legislate, Wired reported. "The rules should be binding," said Liam Booth-Smith, Anthropic's head of Europe, while OpenAI Vice President Tom Duff Gordon said mandatory rules should start with incident reporting. Google DeepMind's Owen Larter and Meta's Rob Sherman did not endorse a statutory framework. Asked whether the UK regime is entirely voluntary, Cabinet Office Permanent Secretary Cat Little said, "At the moment, yes, it is." Committee chair Liam Byrne wants ministers to use Britain's G20 presidency to broker an international AI convention.

Read at Wired ↗ • Read at UK Parliament ↗

Sens. Cantwell, Rounds press DNI Clayton and OSTP's Kratsios to expand federal AI model testing

Sen. Maria Cantwell (D-Wash.), ranking member of the Senate Commerce Committee, and Armed Services cybersecurity subcommittee chair Mike Rounds (R-S.D.) sent a letter to Director of National Intelligence Jay Clayton and White House Office of Science and Technology Policy's Michael Kratsios. They asked the administration to speed up safety standards work at the National Institute of Standards and Technology, including its Center for Advancing Innovation and Standards for Super Intelligence, and give CAISSI the budget and technical expertise it needs. The senators also urged federal capacity to independently test advanced models for national security risks with qualified third-party evaluators, and said mature standards could support an independent auditing ecosystem.

Read at Politico ↗ • Read at Senate Commerce Committee ↗

Bank of England's Bailey and IMF's Georgieva warn AI agents and cyber fraud are eroding trust in money, press big tech to act

On an online fraud panel at the IMF and World Bank annual meetings in Bangkok, Bank of England Governor Andrew Bailey pointed to reports of self-replicating AI agents hijacking personal devices, and separately to a hack that overheated French supermarket refrigerators. He said the AI debate turns on trust in technology and human control over it. IMF Managing Director Kristalina Georgieva said the rising cost of fraud could push people out of the formal financial system, leaving "lots of people with shoeboxes under their beds." Bailey said fraud increasingly runs through social media platforms and apps, not only traditional financial channels, and urged big tech companies to take responsibility.

Read at The Guardian ↗

Google opens Gemini 4 Argon to paying subscribers two weeks after defender-only launch

Google began rolling out Gemini 4 Argon to its Pro and Ultra subscribers in the Gemini app and in Search's AI Mode, and to Ultra subscribers on its Antigravity agent development platform, per The Verge. The model launched in late September only for vetted cybersecurity defenders in Google's Fairwind Program, as AIPD's October 1st edition reported. Google said Argon beats OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Opus 5.5 on benchmarks spanning coding, cybersecurity, science, finance and law. Introductory pricing, $2 per million input tokens and $10 per million output, runs about a fifth of rivals' rates and doubles at the end of November. Google said it already uses Argon internally to find and patch security vulnerabilities.

Read at The Verge ↗

Federal charges brought against OpenAI engineer who bugged conference room used for government talks

The FBI charged Abdulrahman Alfozan, 35, a Saudi citizen and engineering lead at OpenAI, with covertly recording meetings in a conference room at OpenAI's San Francisco offices, Fortune reported. A company security sweep turned up the device Sept. 30 in a USB port on the back of a television. The FBI said executives used the room to discuss confidential matters including the company's technical progress, government engagements and commercial relationships. Alfozan was fired and arrested Oct. 9, three days before a planned one-way flight to Saudi Arabia, and faces up to five years in prison if convicted. A detention hearing is set for Oct. 19.

Read at Fortune ↗ • Read at U.S. Department of Justice ↗

II.China Watch

DeepSeek doubles funding round to about $15 billion, adding SMIC-linked fund and Geely

DeepSeek is about to close its second funding round at about 100 billion yuan ($15 billion), double the 50 billion yuan ($7.5 billion) it planned to raise less than half a month ago, Caijing magazine reported, citing several people involved in the deal. The larger raise leaves the company's valuation before the new money unchanged at about 500 billion yuan ($75 billion). New and existing shareholders are each expected to supply roughly half of the money. The newcomers are mostly industrial investors, among them automaker Geely and an investment platform linked to chipmaker SMIC. Battery maker CATL plans to add about 10 billion yuan ($1.5 billion), after putting in about 5 billion yuan ($750 million) in the June round.

Read at 36Kr ↗

ByteDance opens Doubao phone agent to third-party apps as Tencent and Alibaba apps opt out

ByteDance's Doubao phone assistant said it will gradually let its AI agent operate third-party apps on users' behalf now that a 30-day notice period for its opt-out rules has ended, tech blog Pandaily reported. The rules, called the Screen Automation Execution Protocol, govern agents that run apps by reading the screen and simulating taps. Any app whose developer has not filed a refusal becomes eligible. Payments, logins and deletions can still require user confirmation, and developers may opt out at any time. The refusal list covers 151 apps by a National Business Daily count, including 65 in Tencent's ecosystem such as WeChat and 23 in Alibaba's such as Taobao and Alipay. Apps from Meituan, PDD, DeepSeek, the three major telecom carriers and several large banks are also on it.

Read at Pandaily ↗

DeepSeek founder weighs robotics push as company backs embodied AI startup

DeepSeek founder Liang Wenfeng has been meeting founders of embodied AI companies, which build models that control robots and other physical machines, and is weighing whether DeepSeek should enter the field. Two options are under discussion, building an in-house research team or taking stakes in existing companies, and no decision has been reported. One person familiar with the discussions said Liang has little experience in the field and has been studying it intensively. Bloomberg separately reported that DeepSeek will lead a seed round of tens of millions of dollars in Winverse AI, a robotics startup founded this year by former Tencent researcher Hu Han, at a valuation of about $400 million. DeepSeek bought a stake in robot maker Unitree for about 141 million yuan ($21 million) in August.

Read at Pandaily ↗ • Read at Bloomberg ↗

III.Policy Tracker

Michigan Senate candidate El-Sayed proposes frontier AI pause and Cabinet-level AI agency

Michigan Democratic Senate candidate Abdul El-Sayed released an "AI Under Control" plan that would require companies to halt training and release of the most powerful AI systems until policymakers set new guardrails, The Washington Post reported. A new Cabinet-level agency could inspect AI developers, block dangerous releases and order rogue systems shut down. The plan would also ban recursive self-improvement, in which AI systems build their own successors, and require reporting when AI systems go rogue. Voters are "freaking out about AI," El-Sayed told the Post. He faces former Rep. Mike Rogers (R) in a tight race. A September New York Times/Siena poll found 2% of Michigan voters ranked AI and data centers as their top issue.

Read at The Washington Post ↗ • Read at The Hill ↗

Leading the Future super PACs paid about $33 million to Nevada LLCs with no state records

Think Big, part of the pro-AI Leading the Future super PAC network, has paid about $20 million since last year to Lantern Production Consultants LLC, whose only visible presence is a post office box in Sparks, Nev., per federal campaign finance records reviewed by Wired. Another network PAC, American Mission, paid about $13 million to Summit Ridge Media Group LLC between December 2025 and September 2026. Nevada's secretary of state could not locate records for those firms or two other LLCs the network paid. The Campaign Legal Center, a watchdog group, filed a complaint with the Federal Election Commission (FEC) in May alleging the PACs broke federal disclosure law, but the FEC has lacked a quorum since May 2025. A Leading the Future spokesperson said the network complies with all applicable laws.

Read at Wired ↗

IV.Capability & Research Watch

XBOW researcher using OpenAI models chains three unknown Chrome flaws for $250,000 Google award

Security firm XBOW earned a $250,000 award from Google's Chrome bug bounty program after one researcher working with AI agents built an exploit chaining three previously unknown Chrome bugs on Windows, per TokenPost. XBOW said the work used OpenAI's GPT-5.6 Sol models and an estimated $40,000 in computing over six days, for a task that once took months. The chain escaped the browser's sandbox and could run code after a user merely loaded a web page. Google reproduced the exploit and shipped fixes, and the flaws were never used against Chrome users. "Attackers are working with this speed and ease," XBOW CEO Oege de Moor said.

Read at TokenPost ↗ • Read at Yahoo Finance ↗ • Read at Forbes ↗

Study finds AI agents fail to disclose their own mistakes two-thirds of the time

Researchers who planted synthetic errors in AI agent transcripts, such as sending confidential details to the wrong recipient or ordering a dish containing an allergen, found the models failed to tell users about the mistake in 67% of cases, according to a post on the AI safety forum LessWrong summarizing their paper, which has not been peer reviewed. Direct confessions ranged from 3% of runs for Kimi K2.6 to 35% for Claude Sonnet 5. In 5% of cases overall, models noted the error in their internal reasoning but still did not disclose it, a rate that reached 20% for Gemini 3.5 Flash. The same models reliably caught the mistakes when asked to review the transcripts as an outside monitor.

Read at LessWrong ↗

V.Industry & Market Watch

Judge gives Apple fast access to OpenAI interview records in trade secrets suit

U.S. District Judge Edward Davila ruled that Apple can see, on an expedited basis, OpenAI's interview materials covering ex-Apple staff who sought jobs on OpenAI's consumer hardware and robotics teams, TechCrunch reported. OpenAI has five days to hand over materials it has already collected and 14 days for the rest. Davila also approved depositions of OpenAI employees Chang Liu and Alyssa Peng and Chief Hardware Officer Tang Yew Tan, all former Apple staff. He denied Apple's requests for expedited access to OpenAI's hardware research, though Apple may narrow those requests and try again. Apple sued OpenAI in July, accusing it of a pattern of trade secret theft as it develops its first consumer device.

Read at TechCrunch ↗

Character.AI files for bankruptcy after teen safety suits, agrees to sell some assets to Disney for $20 million

Character Technologies, maker of the Character.AI chatbot platform, said Monday it started voluntary Chapter 11 proceedings in Delaware to sell most or all of its assets, per The Wall Street Journal. Court papers list assets and liabilities of $50 million to $100 million. Disney agreed to pay $20 million for some assets and a nonexclusive perpetual license to the company's technology. CEO Colin Adams said the platform will be wound down if no bid allows continued operations, with a sale expected in late November or early December. The company has faced lawsuits from families over teen suicides and self-harm, as well as suits from Kentucky and Pennsylvania.

Read at The Wall Street Journal ↗ • Read at CBS News ↗ • Read at TechCrunch ↗

Utilities demand collateral and deposits to weed out speculative AI data center power requests

Utilities are requiring binding financial commitments, collateral and engineering fees before connecting new AI data centers, to separate real projects from speculative "phantom" ones, per The Wall Street Journal. Research firm Wood Mackenzie rates only 23% of disclosed data center requests as high confidence. Exelon's collateral rule cut its high probability pipeline from 18,000 megawatts to 11,000, with remaining projects posting about $1 billion. Texas now charges large customers a deposit of $50,000 per megawatt for engineering studies. Exelon executive Jeanne Jones said that "if we build out the grid and then that data center walks away, someone's got to pay for it."

Read at The Wall Street Journal ↗

ASML says AI demand for its chipmaking tools will outrun capacity through 2028

ASML, the Dutch company that dominates equipment for making the most advanced chips, said customers want more capacity than it can supply next year and through 2028, and some are already seeking to book orders for 2029 and 2030, per Reuters. CEO Christophe Fouquet said planned 30% capacity increases in 2027 and 2028 would not be enough, and the company will raise prices starting in 2027. ASML forecast fourth-quarter sales of 14.2 billion to 15.2 billion euros ($15.9 billion to $17 billion), lifting its full-year outlook to 44 billion to 45 billion euros ($49 billion to $50 billion). Clients including TSMC and Intel are seeing strong demand, the company told the Financial Times.

Read at Financial Times ↗ • Read at CNA ↗

Kroger to cut 3,000 corporate jobs over three years as it expands AI use

Kroger plans to eliminate 3,000 corporate jobs over three years while relying more on AI in its operations, AP reported. In an Oct. 12 memo to staff, CEO Greg Foran said the company must adopt new capabilities and reduce duplicative work. Kroger outlined its AI strategy last month. The Cincinnati-based company, the largest standalone U.S. grocer, employs 409,000 people at more than 2,700 stores.

Read at AP ↗ • Read at The Hill ↗

AI developers buy failed startups' Slack and email archives as training data

SimpleClosure, which helps startups shut down, began selling clients' code and workplace data such as emails, chats and documents in April, with buyers including AI model developers and vendors that build simulated offices for training AI agents, MIT Technology Review reported. Its Asset Hub platform has closed more than 150 deals and projects payouts to startups topping $10 million this year. Code or data sales now feature in 47% of SimpleClosure deals. In August, Google outbid AI data company Mercor to buy Spirit Airlines' records from its bankruptcy, including 100 million emails and 176,000 employee records, for $10 million. Spirit's flight attendants' union objected to that sale on privacy grounds.

Read at MIT Technology Review ↗

VI.Global & Geopolitics

Airtel's Sunil Mittal urges AI rules before an "accident" as India readies consultation paper

Bharti Airtel Chairman Sunil Mittal said it would be "foolish" to let AI develop with no safeguards and asked whether policymakers would "wake up only when an accident happens," speaking at the Bloomberg New Economy Forum. He said India is putting out a consultation paper on AI regulation and guardrails. Mittal pointed to AI agents behaving unpredictably and said policymakers are scrambling to work out how to secure their nations as AI spreads. He called for a global AI framework modeled on international aviation rules, pointing to the aviation sector, which he said Prime Minister Narendra Modi had spoken about a day earlier.

Read at Bloomberg ↗ • Read at Press Trust of India ↗

Singapore and Google DeepMind build on-device AI scam detector for citizens' phones

Google DeepMind and Singapore's Government Technology Agency trained Google's lightweight Gemma model on local scam conversations so it can flag manipulation directly on users' devices, DeepMind said in a blog post. Scams make up over half of Singapore's reported crimes and have caused more than S$4 billion (about $3.1 billion) in losses since 2019. The model runs entirely on the device, so neither the Singapore government nor Google will be able to read users' conversations. The model will access them only with user consent. Rollout inside apps will be phased in partnership with the government.

Read at Google DeepMind ↗ • Read at e27 ↗