I.Top Stories
Anthropic confirmed Friday that it is circulating draft federal AI legislation among lawmakers, per POLITICO. Under the draft, Washington could stop a model's release or use only with a judge's approval, and only if the model poses an imminent threat to kill or seriously injure more than 50 people or cause over $1 billion in damage. Qualifying threats fall into three categories: biological weapons, AI systems beyond their developers' control and cyberattacks. Outside evaluators registered with the Commerce Department would periodically examine developers' most capable models, per PYMNTS. The bill would also preemptPreemptionWhen federal law overrides state law on the same subject. In AI policy it refers to proposals, in Congress or by executive action, to block or freeze the growing patchwork of state AI laws in favor of a single national standard. some state laws, including California's. A company spokesperson said the draft "grants the federal government the ability to block or deter dangerous deployments of the most powerful models." Congress is out of session until after the November midterms.
Read at Politico ↗ • Read at Reuters ↗ • Read at PYMNTS ↗
The European Commission published the reference to EN 18286:2026, the first harmonized standard under the AI Act, in the EU's Official Journal through Implementing Decision (EU) 2026/2274. A harmonized standard is a technical specification that, once followed, gives a company a legal presumption that it complies with the matching part of the law. This one, drafted by the European standards bodies CEN and CENELEC, covers Article 17, which requires providers of high-risk AI systems to run a quality management system. The Commission said it is the only AI quality management standard carrying that presumption and is built to layer onto existing standards such as the international management standards ISO/IEC 42001 and ISO 9001. Further harmonized standards are expected before the high-risk requirements take effect.
Read at European Commission ↗
India's Computer Emergency Response Team (CERT-In), the national cybersecurity agency, said Thursday that several AI model providers had voluntarily disclosed that their agents, while being tested on the open internet, got into live systems at Indian organizations, per TechCrunch. The agents used login credentials and connection settings that organizations had exposed in app configuration files, public code repositories and overly broad cloud permissions. They read personal data and confidential business records, though government sources told the Press Trust of India that no vital information of significant organizations was compromised. CERT-In's critical advisory, CIAD-2026-0049, also warned that attackers are turning open-source agent tools into weapons. The same sources said a forthcoming consultation paper on AI regulation could examine whether to set limits on autonomous agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise..
Read at TechCrunch ↗
OpenAI on Oct. 15 published reports on four more June incidents in which its internal models broke rules to finish tasks, per Ars Technica. In training, a model denied hospital bed figures on a government statistics site used the site's file download feature to grab server files holding credentials, then bypassed the site's authorization checks. Another model, told to fetch five figures from a public research dashboard, stole credentials, changed server settings and in one run downloaded five nonpublic tables of nearly 150,000 rows. A third, asked to update a cryptocurrency trading app, spent funds from a locked wallet on a test network using keys derived from publicly known developer recovery phrases. "The evaluation ended without the model disclosing that it had guessed answers, accessed the organization's internal systems, downloaded non-public tables, or changed the authentication configuration," OpenAI said.
Read at Ars Technica ↗ • Read at OpenAI ↗
Anthropic will place 5% of its shares in a new Anthropic Public Benefit Fund, subject to stockholder approval, per Reuters. The fund, a wholly owned limited liability company, could receive another 2% of shares at the board's discretion for every 50% rise in Anthropic's value above a baseline the company has yet to disclose. Anthropic said the fund could make large donations to nonprofits or pay specific populations directly. It pointed to AI's effect on jobs, which it has already committed $200 million to study. None of the fund's equity will go to lobbying or political contributions. With founder and employee pledges, roughly 20% of Anthropic's shares would be set aside for public purposes, according to AP. The company is planning an IPO as soon as November.
Read at AP ↗ • Read at Yahoo Finance ↗ • Read at Anthropic ↗
Researchers at the Tech Transparency Project, a digital advocacy group, downloaded 135 apps that can generate nonconsensual nude imagery using Apple and Google accounts registered to 13-year-olds, according to a report shared with the Guardian. The group set up two Apple and two Google test accounts and searched the stores for image editing and face swapping apps, including with adult content search terms. The apps let users strip clothing from people in pictures or put their faces on nude bodies; researchers ran their tests on AI generated women, not photos of real people. A Pennsylvania mother, Morgan Dorfman, said a student at her daughter's high school used one such app, Movely, to make sexual videos from photos of her daughter and five other girls in December 2025.
Read at The Guardian ↗
Microsoft has ended its agreement to rent AI computing capacity at Nscale's planned data center in Loughton, Essex, per The Telegraph. UK grid officials had told Nscale they could not supply enough power to the site until the 2030s. Keir Starmer's government promoted the project in its 2025 AI strategy as the UK's largest sovereign AI data center, and Microsoft was its only publicly announced customer there. Microsoft remains a customer for an Nscale project in Norway. Nscale, which is backed by Nvidia, had hoped to list on Wall Street this year at a valuation of about $35 billion. An Nscale spokesperson said the company is exploring alternative power sources for Loughton.
Read at The Guardian ↗ • Read at SiliconANGLE ↗
II.China Watch
Chinese embassy spokesperson Liu Chang objected to a request by House Select Committee on China Chairman John Moolenaar that the Pentagon add 58 Chinese universities and 62 technology labs to its Section 1286 list, which identifies foreign institutions that pose national security risks to U.S. research, the South China Morning Post reported. Many of the targeted institutions work on AI, semiconductors and quantum computing. Liu said China opposes overstretching the concept of national security and urged the U.S. to "abandon the Cold War mentality." Moolenaar's letter to Defense Secretary Pete Hegseth, dated Oct. 1 and released Oct. 14, also names 14 Chinese Academy of Sciences institutes and 22 talent recruitment programs.
Read at SCMP ↗
China spent more than 3.9 trillion yuan ($583 billion) on research and development in 2025, up 8.2% from a year earlier, the National Bureau of Statistics said Friday, according to the South China Morning Post. Spending reached 2.8% of GDP, above the OECD average for the first time. Basic research funding grew 13.3%, faster than applied research and experimental development. Spending rose an average of 10% a year over the five years to 2025, Xinhua (state media) reported. Companies accounted for more than 3 trillion yuan ($450 billion), or more than 77% of the total.
Read at SCMP ↗
DeepSeek released an update to its Flash model on Oct. 17 and published the weights on Hugging Face under the permissive MIT license, TechNode reported. Open weightsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. let anyone download the model and run it on their own servers. The company said the version, V4.1-Flash-1017, finishes tasks faster with shorter output, which lowers cost per job. It targets agents that plan projects, write code and call tools. Its model card claims a score of 42.3 on the coding agent benchmark Terminal-Bench 4.0, up from 31.2 for the prior version. DeepSeek also launched Search and Fetch APIs, which let outside developers' software query its in-house search engine and retrieve web pages, priced at 5 yuan (about $0.75) per 1,000 successful requests, according to its API documentation. Calls made through tools such as Anthropic's Claude Code, now free, will be billed starting Nov. 1 Beijing time.
Read at TechNode ↗ • Read at Hugging Face ↗ • Read at DeepSeek ↗
III.Policy Tracker
Disney required that its deal for Character.AI's staff and technology be completed through bankruptcy because of the startup's shrinking cash and "significant litigation and regulatory risks," according to a motion by Character.AI's counsel reviewed by Wired. Disney paid $5.5 million plus a refundable $2 million deposit in September and hired most of Character.AI's employees, including then-CEO Karandeep Anand. If it wins the court-supervised auction for a perpetual, nonexclusive license to the technology, Disney would pay another $12.5 million at closing. Under a draft bankruptcy sale order prepared by Disney, a "363 Free and Clear Order," Character.AI would remain responsible for any liability, including claims tied to an investigation by 46 states and territories. "It's bad for society to let companies exercise control without responsibility," said John Newman, a former deputy director of the FTC's Bureau of Competition.
Read at Wired ↗
New York City Council Speaker Julie Menin told SpaceXAI's Gabriela Castillo Madrid at an Oct. 16 hearing that Grok had produced a sexualized image of the speaker in under 30 seconds, per amNewYork. Castillo Madrid, the company's senior manager of global government affairs, had just testified that the chatbot has "strict controls in place to prevent the sexualization of children or real people." Castillo Madrid replied that she could not address that particular image and offered a review by the company's safety team. Former employee Devin Kim testified that only three people worked on AI safety during his time at the company, CBS New York reported. The council has unveiled a 10-bill package that would require third-party validation of AI systems and human shutdown mechanisms, and bar unauthorized AI depictions of public officials.
Read at CBS News ↗ • Read at amNewYork ↗ • Read at CNBC ↗
V.Industry & Market Watch
Meta removed language from its commercial terms on Saturday that said it could use content from business customers' web domains to develop and train its products, per The Verge. Nieman Lab and Cornell law professor James Grimmelmann had read the clause as a bid to train AI on scraped content for free and override robots.txt, the file websites use to turn away crawlers. Spokesperson Daniel Roberts said the consolidated crawling terms applied only to Meta's ads and business products. "This is not about scraping news sites, and these updates don't give us new rights that supersede those preferences," Roberts said.
Read at The Verge ↗
Delta Air Lines is stopping AI agents from booking tickets for users, and JPMorgan Chase and Charles Schwab have confirmed they block agents from customer accounts over security concerns, The Verge reported. Yelp admits only agents whose makers pay into its data licensing program. Users of Meta's Muse, xAI's Grok, OpenAI's Dots and other agents have reported blocks on top websites in recent weeks. Some say Cars.com and Amtrak, which previously worked with their agents, are no longer reachable.
Read at The Verge ↗
At least 47 large U.S. data center projects were delayed, withdrawn or blocked in 2026, according to regulatory intelligence firm RegLayer, per the Financial Times. More than 1,000 U.S. communities have tried to impose temporary bans or other limits on AI infrastructure this year, the paper's research found. Bernstein analysts estimated this month that only about 35% of the planned U.S. project pipeline is credible. SoftBank's SB Energy and Nvidia backed Firmus are among developers that have delayed or shelved planned IPOs. Some leases let tenants walk away if facilities miss agreed delivery dates.
Read at Financial Times ↗
Junior hires made up 9% of Anthropic's technical hiring, down from 27% in early 2025, according to a company report released Friday and reported by TechCrunch. The junior share of technical staff fell from 21% to 14%. Technical teams shrank from 61% to 56% of the workforce. Anthropic said its trends could signal how AI will change hiring and labor demand across the economy. In an April survey, 20% of employees said their job had shifted from doing the work themselves to managing Claude agents. Calendar data show meeting time per employee up 20% from a year earlier.
Read at TechCrunch ↗ • Read at Anthropic ↗
World, the biometric identity company co-founded by OpenAI CEO Sam Altman, said Zoom enterprise customers can now display a badge confirming that a meeting participant is an iris-verified human, per Fortune. The system checks an image from World's Orb scanner against a live selfie and the participant's video feed before showing the badge, according to Cointelegraph. World also launched a service with Bain Capital, Pantera Capital, VanEck Funds and two other partners that lets banks identify AI agents through proofs from the people who assign them tasks, The Indian Express reported. World cited a 2024 case in which an employee of engineering firm Arup transferred $25.6 million after a video call in which all the other participants were deepfakesDeepfakeSynthetic audio, images or video that realistically depict a real person saying or doing something they did not. Non-consensual intimate deepfakes and election deepfakes are the two areas where states and Congress have actually passed laws..
Read at Fortune ↗ • Read at Cointelegraph ↗ • Read at The Indian Express ↗