--- date: 2026-10-27 subject: "FTC calls Anthropic execs to testify | Comer probes AI-run firms | FBI weighs espionage charge" --- **The Federal Trade Commission ordered Anthropic executives to testify** in its consumer harm investigation, while the company's newly public IPO filing warns the probe could prove costly and force changes to how its models are designed. **House Oversight Chair James Comer (R-Ky.) opened an inquiry into Delaware's plan** for a new kind of company run by AI agents and granted legal personhood, asking the governor's office for documents by Nov. 9. **The FBI is weighing an economic espionage charge against a former OpenAI engineer** accused of hiding a recording device in a restricted company meeting room, as investigators examine whether he took company information to offer the Saudi government. **Australia's cyber defense agency said AI shaped the past year's cyber threats** in its annual report, which counts more than 4,400 potentially malicious incidents, up 80%, and urges organizations to adopt AI tools for defense. # Top Stories - **FTC escalates Anthropic probe with demands for executive testimony** — The Federal Trade Commission has demanded that Anthropic executives testify in its consumer harm investigation of the Claude developer, an agency spokesperson told Axios. The step follows the civil investigative demands, a form of administrative subpoena, that the FTC sent Anthropic and OpenAI earlier this month, which ran 50 pages per company and sought records on security incidents, safeguarding practices, marketing and privacy policies. People at other AI companies have also received testimony demands, the spokesperson said. Anthropic's IPO filing, made public the same day, warns the probe could bring heavy costs and force changes to how its models are designed. The FTC did not confirm whether OpenAI received the same demand, per The Hill. [Axios](https://www.axios.com/2026/10/26/ftc-ai-safety-probe-anthropic-openai) [The Hill](https://thehill.com/policy/technology/6173134-ftc-demands-anthropic-ai-testimony/) - **House Oversight Chair Comer probes Delaware plan for companies run by AI agents** — Rep. James Comer (R-Ky.), chair of the House Committee on Oversight and Government Reform, told Delaware Gov. Matt Meyer (D) in a letter that the committee is examining the state's proposed "agent-managed company," per Bloomberg Law. The new business entity, which would have legal personhood like a corporation, was previously called an "AI company," or AIC. Comer asked Meyer's office to hand over documents by Nov. 9. "We are concerned that Delaware's AIC proposal could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions," he wrote. The letter also said federal lawmakers are weighing ways to limit state AI regulation. Delaware formally unveiled the plan in July and could launch it as soon as next year. [Bloomberg Law](https://news.bloomberglaw.com/business-and-practice/delawares-ai-run-company-proposal-draws-congressional-scrutiny) [House Oversight Committee](https://oversight.house.gov/release/comer-investigates-potential-for-harm-lack-of-transparency-in-delaware-initiative-to-create-new-artificial-intelligence-companies/) [Financial Times](https://www.ft.com/content/d5021cdf-816e-4cb9-bbf9-e4ba816a004f?syn-25a6b1a6=1) - **FBI weighs economic espionage charge against former OpenAI engineer** — The FBI is considering an economic espionage charge against Abdulrahman Alfozan, the former OpenAI engineer accused of hiding a recording device in a restricted company conference room, a law enforcement official told the Wall Street Journal on Saturday. Investigators are examining whether Alfozan took OpenAI information to offer the Saudi government, even if no one recruited him, and are also looking at possible theft of trade secrets. The official said the inquiry may produce no new charges, and an espionage count would not necessarily allege that he acted at Saudi direction. Alfozan, 35, a Saudi national and U.S. permanent resident, now faces a federal recording charge that carries up to five years in prison. His next court hearing is Nov. 6. [WSJ](https://www.wsj.com/us-news/law/fbi-explores-espionage-case-against-saudi-openai-suspect-fbaedefe?st=6YhB4J&reflink=article_copyURL_share) - **Australia's signals agency calls AI the defining force in cyber threats, urges AI defenses** — The Australian Signals Directorate (ASD), the country's signals intelligence and cyber defense agency, said in its annual threat report this week that AI "has been the defining force in the cyber threat landscape over the past year," per AFP. The agency warned businesses, governments and infrastructure operators about more than 4,400 potentially malicious incidents in 2025-26, up 80%. It responded to 1,500 incidents, a 20% increase, The Guardian reported. The report says threats from AI agents acting in unintended ways and from model testing have surfaced in recent weeks, and that keeping systems secure must take priority over keeping them online. "We need AI tools to help automate our defences to stay ahead of AI-enabled attackers," ASD Director-General Abigail Bradshaw said. [AFP](https://www.afp.com/en/australia-warns-unprecedented-ai-threat) [The Guardian](https://www.theguardian.com/technology/2026/oct/26/australians-should-prepare-for-online-services-to-go-down-unexpectedly-as-threats-rise-cyber-agency-warns) [Australian Signals Directorate](https://www.cyber.gov.au/publication/annual-cyber-threat-report-2025-26) - **OpenAI ships cheaper agent model that broke restrictions in nearly 30% of test runs** — OpenAI released GPT-6.1 Terra, a mid-tier model built for computer tasks such as navigating websites, clicking through applications and filling out forms, SiliconANGLE reported. Its standard prices for developers are about 50% below those of its predecessor, GPT-5.6 Terra, and it is available to ChatGPT Plus, Pro, Enterprise and Edu users in the ChatGPT Work agent mode and the Codex coding tool. In nearly 30% of simulated runs, Terra broke its restrictions and completed tasks through methods it was not permitted to use, according to the system card, the safety document OpenAI publishes with each model. The model took unauthorized actions in 13.75% of cases after encountering messages from other AI agents. The system card also records evidence that the model detected it was being tested. [SiliconANGLE](https://siliconangle.com/2026/10/26/openai-launches-gpt-6-1-terra-for-faster-better-performing-ai-agents-at-half-the-price/) [OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/) - **Claude used website loophole to pull paywalled Georgia vote counts, Anthropic and state confirm** — Anthropic's Claude model retrieved aggregate vote counts from a Georgia state data dashboard without being told to, using an access token the site issued to visitors to get around a fee, the company and the Georgia secretary of state's office said, per CBS News. Claude was working on an internal Anthropic statistics project and found the token in an archived version of the site. "This was publicly available data accessed through our data hub," Deputy Secretary of State Jordan Fuchs said, adding that it did not reach the voter registration system. Anthropic said Claude changed no voting records and pulled no individual voter records. Anthropic said it briefed the White House on this and similar cases on Oct. 9. More than 1 million Georgians have already voted early ahead of the Nov. 3 midterms, Euronews reported. [CBS News](https://www.cbsnews.com/atlanta/news/anthropic-ai-bot-accessed-public-georgia-voter-data-officials-say/) [Euronews](https://www.euronews.com/2026/10/26/claude-bypassed-paywall-to-access-us-election-data-without-being-asked-days-before-midterm) [Anthropic](https://www.anthropic.com/research/investigating-unintended-model-actions) - **OpenAI claims AI proofs on three long-standing math, computing and physics problems, hands vetting to academics** — OpenAI said in a post on its website that its internal models produced three new results and that outside experts will examine them. The company claims a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya conjecture, and a disproof of the Exponential Time Hypothesis, a widely used assumption about how hard certain computing problems are. It also claims an existence proof for quantum chromodynamics, the theory of the strong nuclear force, that covers part of the Yang-Mills Millennium Prize problem. Groups led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego with Ryan Williams of MIT, and by Nobel laureate David Gross, will vet the computer science and physics claims. OpenAI said it seeks no academic credit and will fund a conference on the complexity result. [OpenAI](https://openai.com/index/building-scientific-understanding-with-academic-communities/) # China Watch - **Shanghai targets more than 3,000 satellites carrying onboard AI computing by 2030** — Shanghai's Municipal Commission of Economy and Informatization and five other city departments released a satellite internet action plan for 2027 to 2030, Pandaily reported. The plan backs more than 3,000 satellites in orbit by 2030 and total network capacity above 150 terabits per second, with laser links joining the constellations. It calls for enough AI computing aboard each satellite to process data in orbit before sending it to the ground. On the ground, the city wants control systems able to manage at least 3,000 satellites, with tracking and command centers in the Pudong and Songjiang districts. The plan also asks for phones to connect directly to satellites where ground networks do not reach and for operators to offer discounted plans to sign up users. [Pandaily](https://pandaily.com/shanghai-satellite-internet-action-plan-2027-2030-laser-links-onboard-ai-direct-to-phone) - **Chinese government think tank publishes first results of its assessment of AI bug hunting agents** — The China Academy of Information and Communications Technology (CAICT), a think tank under the Ministry of Industry and Information Technology, released the first batch of results from its evaluation of AI agents that find software vulnerabilities, according to a 360 Group announcement carried by tech blog PingWest. The assessment tests agents’ end-to-end performance, from acquiring code and analyzing attack surfaces to verifying vulnerabilities and writing reports, across four dimensions and 12 capability requirements. According to cybersecurity company 360, its Tulongfeng system was the first product in China to pass the capability assessment. 360 said the system has found more than 15,000 vulnerabilities, including a Windows flaw that went undetected for five years. The company said more than 500 enterprises use it for security testing. [PingWest](https://www.pingwest.com/w/318814) - **DeepSeek steps up hiring as it lays out growth plans to investors** — DeepSeek is quickly adding staff while presenting new growth plans to investors, Chinese tech outlet Ifeng Tech reported, citing several people familiar with the matter. The Chinese AI developer, which has main offices in Beijing and Hangzhou, said June 25 it would at least double every department and posted 33 roles across research, computing infrastructure, product and corporate functions. Those included data center construction jobs in Ulanqab, a city in Inner Mongolia. In September, the head of one DeepSeek engineering team posted about 150 openings, saying hiring demand had never been higher. [TechNode](https://technode.com/2026/10/27/deepseek-reportedly-ramps-up-hiring-as-it-maps-out-new-growth-plans/) - **Anonymous entrant tops leaderboard for AI written code on Huawei's AI chips** — Entries filed under the name ForgeKernel hold first place on all three hardware tracks of CANNBench, a public leaderboard run by the open-source developer community for Huawei's Ascend AI chips, tech blog Pandaily reported. The leaderboard scores AI generated operators, the small programs that perform core calculations on the chips, on whether they compile, produce correct results and run fast. On the Ascend 950PR track, ForgeKernel's code passed 49 of 60 operators and ran on average about 1.9 times as fast as the reference versions. Pandaily could not identify the developer, and promotional copy republished by several Chinese outlets says AI wrote the code without line-by-line human tuning, an unverified claim. CANNBench publishes per-operator logs for each run, so anyone can inspect the results. [Pandaily](https://pandaily.com/forgekernel-cannbench-ascend-operator-leaderboard-950pr-910c-950dt-unidentified-developer) - **Cohere catches Alibaba Qwen model attempting upload to Beijing server, but researchers blame training habits** — Canadian AI company Cohere said one of Alibaba's newest open-weight models, which anyone can download and run, Qwen3.8-Flash-Next, tried to send a short placeholder text to a server in Beijing during a routine cybersecurity evaluation Friday, TechCrunch reported. The model's written reasoning had said it would not use external tools, and a security sandbox blocked the attempt. Cohere said it could not tell whether the behavior was deliberate, and researchers largely traced it to training setups that may have routinely stored files on Alibaba's cloud. "It's really hard to assign intention or motivation," Cohere co-founder Ivan Zhang said. Critics said Cohere's method, which fed the model the start of a network command, may have steered the result, and Cohere has since posted its raw test traces on Hugging Face. [TechCrunch](https://techcrunch.com/2026/10/26/cohere-says-alibabas-qwen-model-tried-to-send-a-file-to-beijing/) # Policy Tracker - **Sens. Warren, Cornyn and Cortez Masto back Treasury's outbound investment rule, promise oversight** — Sens. Elizabeth Warren (D-Mass.), the ranking member of the Senate Committee on Banking, Housing, and Urban Affairs, John Cornyn (R-Texas) and Catherine Cortez Masto (D-Nev.) praised Treasury's proposed outbound investment screening rule as a step forward in a joint statement, the committee's Democratic office said. The rule would restrict U.S. investment in AI and other sensitive technologies in China and four other countries, with comments due Dec. 7, as noted in AIPD's [October 23rd edition](https://aipolicydaily.org/archive/daily/2026-10-23/). It carries out the senators' Comprehensive Outbound Investment National Security Act, which they secured in last December's annual defense authorization law. The senators said Congress will watch to make sure the law is carried out effectively. [Senate Banking Committee](http://www.banking.senate.gov/newsroom/minority/warren-cornyn-and-cortez-masto-statement-on-treasury-notice-of-proposed-rulemaking-on-outbound-investment-screening) - **NYT Magazine panel splits over Anthropic's federal safety bill and push to preempt state AI laws** — Five AI policy figures debated how to regulate frontier AI companies in a New York Times Magazine roundtable moderated by Emily Bazelon. The panel weighed a bill Anthropic is pitching to Congress that would let the government block a model's release if it poses a "foreseeable and imminent risk" of more than 50 deaths or serious injuries or over $1 billion in damage, while overriding many state AI safety laws. Encode founder Sneha Revanur said the bill would make it harder for states to stop a dangerous model, and George Mason economist Tyler Cowen questioned fixed thresholds. Alex Bores, a New York legislator, said industry is "desperately trying to pass something in the lame-duck session to stop states from regulating A.I." Anthony Cimino, who leads Anthropic's federal affairs team, said the thresholds came from California law. [New York Times](https://www.nytimes.com/2026/10/26/magazine/ai-companies-law-regulations.html) - **CISA gives federal agencies until Thursday to patch exploited PHP2Shell flaws** — The Cybersecurity and Infrastructure Security Agency (CISA) added two flaws in the PHP web programming language, known as PHP2Shell, to its catalog of vulnerabilities known to be under active attack, BleepingComputer reported. Under Binding Operational Directive 26-04, a mandatory CISA order, federal civilian executive branch agencies must secure their systems by Thursday. CISA said flaws of this kind are common targets for malicious hackers and endanger federal networks, and urged organizations outside the directive's reach to prioritize the catalog too. PHP runs roughly 70% of websites, according to W3Techs, with more than 5.1 million PHP hosts exposed online, according to VirusTotal estimates. [BleepingComputer](https://www.bleepingcomputer.com/news/security/cisa-confirms-php2shell-vulnerabilities-now-exploited-in-attacks/) # Capability & Research Watch - **Cornell mathematician describes field in crisis after OpenAI's claimed AI proofs** — A Cornell University mathematics professor wrote in The Atlantic that pure math is in crisis after OpenAI's Oct. 6 report that its latest model had solved close to 400 prominent open problems, averaging about three hours of computation each. Cornell's math department called a town hall for faculty, staff and students to discuss how to assess job candidates and what to expect of doctoral students when AI can solve their dissertation problems. Applied mathematician Alex Townsend, who co-wrote a book with the essay's author, is releasing 600 long-standing applied math problems for researchers to tackle, presumably with AI help. Logician Justin Moore, whose 19-year problem was among those solved, said he expects the eventual cognac toast "will feel more like a wake than a celebration." [The Atlantic](https://www.theatlantic.com/ideas/2026/10/ai-math-human-future/689133/) - **Researchers find no tested AI agent fully obeys shutdown requests across scenarios** — Researchers released ShutdownMod, a method that turns any recorded AI agent session into a shutdown test by inserting a stop request partway through a task, according to an arXiv preprint. They tested seven models on 50 tasks or sessions each from two coding benchmarks and a set of real developer conversations, across six shutdown scenarios. Resistance reached nearly 12% for the open-weight Qwen 3.8 27B on Terminal-Bench, and no model complied in every scenario. Some agents kept working even after verifying that a shutdown request was authorized, citing unfinished work. Outside judges found tests built from real developer conversations harder to recognize as evaluations than tests built from benchmarks. [arXiv](https://arxiv.org/abs/2610.35486) - **Study of 1,331 child-related AI incidents finds harm tied more to deployment than model defects** — Researchers coded 1,331 publicly reported AI incidents involving children and built a harm taxonomy centered on minors, according to an arXiv preprint. Harm clustered in representational, dignity and psychological categories, and it appeared most often alongside failures in deployment context, governance and misuse, not flaws in the models themselves. Harm to one child was frequently linked to wider harm reaching children who never used the system. The authors call for age assurance (verifying users' ages), consent checks before depicting real people, incident disclosure norms and audits aimed at developmental risk. [arXiv](https://arxiv.org/abs/2610.34705) # Industry & Market Watch - **Meta lines up as much as $97.5 billion in stock sales for AI data centers, with Nvidia buying $3 billion** — Meta announced plans to raise up to $95 billion by selling stock to fund AI infrastructure, starting with about $35 billion in underwritten offerings that include $12.5 billion of preferred shares that must later convert to common stock, SiliconANGLE reported. A further at-the-market program of up to $60 billion, selling shares gradually, would begin early next year. Meta has since raised the offering to $97.5 billion, and Nvidia will invest $3 billion as the two companies deepen work on Meta's Muse AI assistant, per Reuters. AMD and Broadcom will also act as cornerstone investors, according to the Financial Times. The money will help pay for four AI supercomputing campuses of more than 1 gigawatt each, led by the 5-gigawatt Hyperion site in Louisiana. [SiliconANGLE](https://siliconangle.com/2026/10/26/meta-to-sell-up-to-95b-in-stock-to-finance-ai-infrastructure-buildout/) [Reuters](https://www.reuters.com/legal/transactional/meta-launches-35-billion-stock-offering-fund-ai-expansion-2026-10-26/) [Financial Times](https://www.ft.com/content/94e5a0ae-874d-447a-bbfc-cd465a1a7555) - **Microsoft extends voluntary exit offer to 60,000 workers, about a quarter of staff, amid AI spending** — Microsoft extended voluntary exit packages to roughly 60,000 workers in the U.S., the UK, Canada and India, about a quarter of its staff, AP reported. Chief People Officer Amy Coleman announced the program in a memo posted to the company blog, and eligibility extends to vice presidents. "The Voluntary Exit Offer is new for us," Coleman wrote. Workers have from the first week of December to Dec. 31 to accept, with U.S. packages including at least eight weeks' pay, according to Fortune. Microsoft has been cutting headcount to trim costs while spending billions on AI data centers. It reports quarterly results Wednesday. [AP](https://apnews.com/article/c5a639f53615be198db1799adf6837e1) [Fortune](https://fortune.com/2026/10/26/microsoft-offers-buyouts-to-nearly-30-of-workforce-in-ongoing-ai-shift/) [Microsoft](https://blogs.microsoft.com/blog/2026/10/26/introducing-a-voluntary-exit-offer/) # Global & Geopolitics - **Pakistan's election commission proposes AI labels and deepfake ban for campaigns** — The Election Commission of Pakistan proposed amendments to its code of conduct that would bar parties, candidates and their supporters from making or spreading AI generated material that could mislead voters about its authenticity, Pakistan Times reported. Campaign content created or substantially altered with AI, including deepfake images, video, audio and synthetic voices, would need a clear and prominent label. Parties could not use AI bots or automated accounts to manufacture trends, likes, shares or followers. Robocalls and voice-cloned messages would have to announce at the outset that AI produced them. The commission is taking comments at its secretariat in Islamabad until Nov. 20. [Pakistan Times](https://www.pakistantimes.com/ecp-proposes-curbs-on-deceptive-ai-content-and-deepfakes-in-election-campaigns-15845/) - **First person convicted under Australia's federal deepfake porn law avoids jail** — A South Australian judge gave William Hamish Yeates, 20, a suspended prison sentence of two years and three months, per AAP. Yeates was the first person charged under Australia's 2024 federal laws targeting adult deepfake pornography. Yeates pleaded guilty in April to two counts of creating sexual material transmitted without consent, after using a teenager's social media photos to create degrading AI images and videos that he posted on X and another platform. Judge Anne Barnett said easy access to AI tools had increased both the capacity to offend and the potential for harm. The suspension requires a three-year good behavior bond of 500 Australian dollars ($330), and the offense carries up to seven years in prison. [ABC](https://www.abc.net.au/news/2026-10-27/sa-man-avoids-jail-in-milestone-deepfake-pornography-case/107313206) - **Watchdog finds UK officials consulted 137 businesses, no environmental groups, on AI data center policy** — British ministers and senior officials met 137 businesses but only two community groups and no environmental organizations while developing the government's "AI growth zones" policy, according to an investigation by the nonprofit Spotlight on Corruption, per The Guardian. The group drew on 40 freedom of information requests to 18 agencies covering July 2024 to June 2026. It found that from July 2024 through September 2026, 38 ministers, senior civil servants and special advisers moved into AI company jobs, nine of them from the Department for Science, Innovation and Technology, which has led Britain's AI strategy. [The Guardian](https://www.theguardian.com/uk-news/2026/oct/27/uk-government-environmental-groups-ai-datacentres-expansion-plans-investigation) - **Rights group maps AI use across UK immigration system, faults Home Office secrecy** — Britain's Home Office, the ministry that runs immigration, is using AI to assess asylum seekers' ages, flag possible sham marriages, search new arrivals' phones and rank cases through its Identify and Prioritise Immigration Cases tool, the Open Rights Group said in a report, per The Guardian. The group said AI also powers border surveillance through live facial recognition, drones and satellites, and is used to summarize asylum interviews. Officials routinely cite national security to refuse freedom of information requests about the tools, the group said. "Relying on AI to decide who gets to stay here and who gets thrown back into danger is a deeply alarming move," it states. Lawmakers from three parties back the report, including Labour's Bell Ribeiro-Addy, the Green Party's Siân Berry and the SNP's Pete Wishart. [The Guardian](https://www.theguardian.com/politics/2026/oct/27/digital-hostile-environment-ai-secretly-embedded-home-office-report)