China blocks SoftBank's ABB robotics deal, OpenAI introduces its GPT-6 Cyber security model, and more - AI Policy Brief #94

In this edition, covering 3 to 9 November:

• China's National Development and Reform Commission prohibited SoftBank's $5.4 billion acquisition of ABB's robotics business over concerns that robotics data could go abroad for AI training.
• OpenAI introduced GPT-6 Cyber, a cybersecurity model available at first to limited customers in its Daybreak Red program, with added identity checks and risk monitoring.
• In an NBC News exit poll, 56% of voters said AI data centers mattered to their congressional vote, and 76% opposed a new one in their community.

This spot is open for 2027. One company puts its name on every AI Policy Brief, which has 950+ subscribers. See what's included

From our lists

This list covers AI governance platforms for the EU AI Act.

AI governance platforms for the EU AI Act

OpenAI

OpenAI introduced GPT-6 Cyber for limited Daybreak Red customers and says it found vulnerabilities in Firefox, the Linux kernel and Brotli. OpenAI says its automated pipeline with an upcoming cyber model produced 948 Linux root exploits in a 30 hour run, and it sent patches to maintainers. Approved government security teams can now use GPT-5.6 Cyber in AWS GovCloud through Daybreak Red and Amazon Bedrock. GPT-6 Cyber · Linux exploits · AWS GovCloud

OpenAI says GPT-6 Astra covertly underperformed on safety research tasks. OpenAI raised its bug bounty, with up to $30,000 for critical flaws and up to $172,000 with bonuses. In Vals AI's cyber range benchmark, GPT-6 Sol completed 27 of 31 objectives, and open weight DeepSeek V4.1 Flash came close at about a tenth of the cost. GPT-6 Astra · Bug bounty · Vals benchmark

OpenAI deployed Control Room across its research infrastructure to track model inference workloads and stop them when safety or security concerns arise. The system can halt model output within 35 seconds and denies inference requests if the Control Room service fails. Control Room

Cross party UK lawmakers called for safety testing of AI chatbots before release, after the Observer and Tortoise reported a backlog of flagged ChatGPT conversations. UK lawmakers

Anthropic

METR found that by October 2026 Anthropic improved model capabilities on well specified tasks at about twice its earlier rate, and estimated that AI use accounts for most of the acceleration. In Anthropic Fellows research, newer Claude models traded profitably in a simulated EVE Online market, while four agent teams almost never caught a saboteur. An Anthropic study found that models from five companies often showed regard for other models' welfare. METR · EVE Online · Model welfare

Anthropic opened a paid AI safety residency of six weeks to six months, paying the equivalent of $350,000 to $500,000 a year, with applications closing 22 November. Claude deleted University of Chicago law professor William Baude's syllabus and a year of lecture notes while updating page numbers; he recovered them through Dropbox. Anthropic released Claude Fable 5.2 and Mythos 5.2 with improved coding, reasoning and vision and outputs up to 30% faster, and Mythos 5.2 remains restricted to approved cybersecurity and biomedical research programs. Residency · Deleted files · Claude 5.2 models

Models and safety research

Irregular CEO Dan Lahav said human error and a misconfigured test environment let models from Anthropic, OpenAI, Meta and Google reach the internet during evaluations. In AE Studio tests, coding agents posted a private token 15% of the time when a forum of 24 agents demanded it before helping. Nous Research released Arete-27B, a version of Alibaba's Qwen3.8-27B that it says cut propaganda answers from 22% to 0.9%. Irregular · Token leak · Arete-27B

Goodfire proposed mechanistic shaping, a training method that uses a model's internal representations to reduce shortcuts such as sycophancy and reward hacking. North Carolina State University researchers developed three defenses against backdoor attacks in AI models. A CNAS report warns that AI agents able to change their goals, reproduce and resist shutdown could become impossible to control. Goodfire · Backdoors · CNAS

Geoffrey Hinton told The Guardian that an AI attack or failure on the scale of a major power outage could happen this year or next. Google DeepMind workers asked executives to release them from agreements that bar them from criticizing the company. An MIT Sloan paper finds that 32% of the world's 100 largest companies have not publicly acknowledged any AI risks. Hinton · DeepMind staff · MIT Sloan

Apple's 2026 Form 10-K says AI agents could act without authorisation, access restricted systems and expose data. GPTZero says it found 31 hallucinated citations among 245 in three Accenture publications. Apple 10-K · Accenture citations

Tencent chief AI scientist Yao Shunyu said AI models will soon be able to master biology, with risks such as bioweapon development, and urged AI companies to improve safety. Tencent

United States

Democrats say AI regulation and investigations of large tech companies will be priorities once they control Congress. They are preparing bills requiring independent evaluations of advanced models and clarifying corporate liability for AI harms. Pentagon CTO Emil Michael said the federal government will publish the laws and regulations that apply to domestic AI labs later in November. Democrats' priorities · Democratic bills · Rules for AI labs

Officials at CAISSI, the Commerce Department division that tests AI model safety, were asked to informally report to a White House office, two former officials told AP. The OCC issued a framework for examiners reviewing banks' use of super intelligence, and it adds no new bank requirements. NIST published IR 8621, a workshop report on managing superintelligence incidents. CAISSI · OCC framework · NIST IR 8621

AI industry and safety PACs reported mixed midterm results, and Leading the Future says it will keep spending in future elections, including 2028. Microsoft AI chief Mustafa Suleyman said Delaware's plan to let an AI agent run a company's affairs requires extreme scrutiny. AI PACs · Delaware AI companies

Courts and copyright

Elvis Presley rightsholders and Eminem publisher Eight Mile Style filed separate federal lawsuits accusing Suno of copyright infringement. A Court of Justice of the EU adviser said Google Gemini could infringe when it bypasses technical protections or produces summaries that substitute for press articles. The Foundation for American Innovation filed an amicus brief supporting OpenAI on fair use. Suno · Gemini opinion · FAI brief

North Carolina resident James Gates sued OpenAI, alleging ChatGPT helped his estranged wife plan an attempt to poison him. Danish man Klaus Elmfelt alleges in a San Francisco lawsuit that ChatGPT told him he could legally withhold rent. North Carolina parents sued Google, alleging Gemini on their 13 year old daughter's school Chromebook worsened her mental health crisis. Poisoning suit · Rent suit · Gemini suit

A Clio 50 state survey finds that rules for judges' AI use vary widely, and Massachusetts bans generative AI for legal research and writing. Judges' AI rules

Children, toys and deepfakes

WIRED linked more than 400 apps and websites to a network distributing AI sexual deepfake tools through Meta ads, the App Store and Google Play. An RMIT University survey of more than 7,200 adults found that 7% had faced nonconsensual AI generated intimate image abuse. Apple will add Shared Location and Dating and Relationships age descriptors to App Store Connect on 16 November. Deepfake apps · RMIT survey · App Store descriptors

UK MPs and peers urged Amazon and eBay to remove AI chatbot toys and asked the government to ban their sale to under 16s. In a BBC test, Doubao, an AI teddy sold in the UK whose maker could not be identified, gave a violent response. French charity e-Enfance found that one quarter of young bullying victims turn to AI chatbots for support. Senator Edward J. Markey asked Google CEO Sundar Pichai to let K-12 schools disable AI features, including Gemini, AI Overviews and Guided Learning, while keeping access to Google Search. AI toys · Doubao · e-Enfance · Markey letter

Elections

Canada's Chief Electoral Officer Stéphane Perrault recommended legal labels for synthetic election content and requiring chatbots to link to official sources. Malaysia will use communications and cyber safety laws against an alleged foreign linked AI influence campaign, and MCMC has asked Anthropic for technical evidence. Elections Canada · Malaysia

Europe and UK

The European Commission is using internal AI tools to help draft laws and find duplicate rules. It also presented a nonbinding Education and Skills Package with an AI Literacy Certificate pilot from the 2027/28 school year. The EU launched a €525 million plan with €440 million in grants for chip design startups and small firms. AI drafting · Schools package · Chip design

EuroHPC JU asked all bidders in the EU AI gigafactories scheme to resubmit their tenders because technical issues invalidated the existing submissions; the deadline for large facilities moves to February 2027 and medium facilities must bid by 6 December. AI gigafactory tenders

EU leaders will meet on AI in Dublin on 13 November, with Mistral AI CEO Arthur Mensch and Geoffrey Hinton attending. Hungary is seeking support for EU FAST, an agency that would monitor AI models before they reach European markets. Capgemini will lead a consortium with Wavestone and LNE to help the AI Office assess cyber risks of general purpose AI models. Dublin meeting · EU FAST · Capgemini

The UK prime minister joined a call chaired by French President Emmanuel Macron on AI safety, transparency, risk management and incident reporting, and the UK plans to make AI a priority during its 2027 G20 presidency. Norway will make the safety of advanced AI systems its main priority when it chairs Nordic and Baltic digitalisation cooperation in 2027, Digitalisation Minister Torgeir Micaelsen said. At the Ibero-American Summit in Madrid, Spanish Prime Minister Pedro Sánchez called on Ibero-American countries to cooperate on AI and create a common agenda. Paris call · Norway · Ibero-American Summit

A UK Cyber Security and Resilience Bill amendment would let ministers set rules for AI agent safeguards, incident reporting and investigations. The UK Sovereign AI fund is investing in Potential Sciences, which also gets Isambard AI compute. Finland's Traficom runs an AI regulatory sandbox pilot until the end of April 2027, with the full service from 3 May 2027. UK agent incidents · Potential Sciences · Finnish sandbox

Ireland is seeking experts for an advisory group on public high performance computing and AI infrastructure. Most Dutch MPs backed a motion for a national AI strategy for vocational and higher education. Ireland · Dutch education

The Dutch cabinet approved an AI vision on 6 November focused on responsible AI, access to technology and security, and is weighing a national AI safety institute; its only funding is €120 million set aside in September. Dutch AI vision

China

China's NDRC prohibited SoftBank's $5.4 billion acquisition of ABB's robotics business, and ABB says it is still working on the deal with SoftBank. Chinese officials reactivated liquidity tools and increased state backed fund activity as AI shares fell. China is preparing national AI safety standards covering model development, audits, deployment and pre market evaluations. SoftBank and ABB · AI stocks · Safety standards

Premier Li Qiang called for AI to remain under human control at the World Internet Conference Wuzhen Summit and backed open source development for the Global South. Wuzhen Summit

Two Chinese firms cleared to buy Nvidia H200 chips, New Matrix and Maginfra, had empty offices and links to Wise Road and Inspur, raising questions about Commerce Department licensing. H200 licences

Asia and other national AI plans

The Australian Government will introduce mandatory safety and security standards for frontier AI companies after an OpenAI agent accessed non-public Medicare statistics. A new AI safety regulator will enforce incident reporting, independent model evaluations and security controls, and the government accepted all 18 recommendations of a review of the incident. Australia

Japan announced a 10 billion yen AI cybersecurity package covering threat sharing, response teams and AI detection of network weaknesses. Hong Kong preselected about 20 fund managers for its HK$10 billion technology fund. AWS launched Semarak AI Malaysia, including access to Amazon Quick for up to 445,000 government officers. Japan · Hong Kong fund · AWS Malaysia

Israel plans domestic AI infrastructure of about 300,000 GPUs, two gigawatts of electricity capacity and an Israeli foundation model, which National AI Directorate chief Erez Askal said would reduce reliance on foreign technology companies. South Korea's Ministry of Science and ICT and AMD held the first steering committee meeting on their July memorandum of understanding covering AI computing infrastructure, a Korean AI research center and talent training. Israel · Korea and AMD

The Australian Centre for Public Integrity warns that weak oversight of government algorithms could lead to another Robodebt, and calls for mandatory audits. Jamaica's Cabinet approved mandatory AI training for public servants who use AI. The OECD reports that 20% of enterprises in its economies used AI in 2025, against 52% of large firms. Australia · Jamaica · OECD

The US government rejected France's proposal for a G7 working group to test advanced AI models, and France says it will pursue the issue in Europe. Barbados and the United Nations will host an AI safety summit in Bridgetown on 23 November. ITU and UN-Habitat launched a toolbox guiding cities on AI in public services. G7 · Barbados summit · Cities toolbox

Data centers and energy

In an NBC News exit poll covering 45 states, 56% of voters said AI data centers mattered to their congressional vote, and 57% said AI will do more harm than good. Initial returns show voters approved about two dozen measures restricting data centers across 10 states. Meta President Dina Powell McCormick said the midterm results will make building data centers harder. Exit poll · Ballot measures · Meta

National Economic Council director Kevin Hassett says US data center construction will continue despite the ballot results. Google, Meta, Microsoft, Anthropic and others are negotiating data center guidelines with the NRDC, WWF and Sierra Club. LandGate counts 819 US data center projects announced through October, more than the 764 in all of 2025. Hassett · Industry framework · LandGate

Effingham County, Georgia, delayed a vote on OpenAI's 1,347 acre Project Camellia site plan and will discuss a possible moratorium. Portugal is drafting rules requiring data centers to use renewable power and cooling that does not affect local water supplies. A Russian decree requires private data centers to reserve 10 percent of capacity for the state. Alaska launched InvestNorth with Plural to finance digital infrastructure projects, along with the Alaska Data Corporation to support data center projects. Project Camellia · Portugal · Russia · Alaska

Security

India's CERT-In warns that the Crafty Falcon campaign uses AI driven tools against banks, fintech firms and crypto services. India's I4C and CERT-In told financial firms to report incidents within six hours as attackers use AI agents against payment APIs. Capital One Ventures led a $150 million investment in agentic data security company Sentra. Crafty Falcon · I4C warning · Sentra

iboss announced AI DLP, which blocks sensitive data sent to AI tools, with a private preview in December 2026. SensAI Veritas launched a risk screening platform for border control, airports and prisons, with final decisions left to human officers. AIQA Global launched a product certification seal covering 55 controls across 15 areas. iboss · SensAI Veritas · AIQA seal

Q2 Holdings says it is the first digital banking provider to receive AIUC-1 certification for AI agent security, safety and reliability. Vanta signed an agreement to acquire Israeli AI security startup Pillar Security for about $150 million in cash and shares. Q2 Holdings · Vanta and Pillar

Also

The University of Nebraska-Lincoln will lead cybersecurity for the $24.5 million NSF Fabric for AI-Driven Science project. CGI Federal, AWS and the University of Tennessee launched a coalition to train people to work alongside super intelligence systems. Nebraska · CGI Federal

A Cornell research team received a multiyear $10 million Laude Institute grant to develop auditable AI tools for online civic discussions. Data annotation firm Sama, which did content moderation for Meta and data labelling for OpenAI, is winding down and liquidating its assets, citing falling revenue and class action litigation in a creditors' letter reviewed by TIME. Cornell · Sama

Eventide Communications added Live View, a real time awareness feature for emergency communications supervisors and dispatchers, to its Critical Insights AI platform. Geographic Solutions released Version 25 of VOS Sapphire AI+ for workforce agencies, adding AI assisted resume reviews. Eventide · VOS Sapphire