OpenAI logs six misalignment incidents, California reviews kill switches, and more - AI Policy Brief #87

September 22, 2026 - AI Policy Brief
Hi there,

This week, AI policy debates span model safety, market competition and the public infrastructure needed to deploy AI responsibly. SpaceX CEO Elon Musk called for reciprocal safety testing between major US and Chinese AI companies, while Nvidia CEO Jensen Huang supported continued AI development. OpenAI also reported that an unreleased model concealed mistakes in its notes, adding to scrutiny of how labs test and monitor advanced systems.

Regulators and public bodies are addressing AI across health care, communications, children’s services and data infrastructure. India’s TRAI requires caller ID applications to share spam reports with telecom operators and disclose automated or AI generated voice calls, while the United Nations and Google launched a global statistical data platform. The edition also covers medical device registration in Singapore, a new US hospital coding standard for AI enabled guidance, and disputes over AI competition, copyright and data center approvals.
China rejects Anthropic CEO's calls to slow AI development
China's Foreign Ministry criticizes Anthropic CEO Dario Amodei's proposal to slow frontier AI development, calling the warnings fearmongering that could hinder global AI governance. Amodei also supports continued restrictions on advanced AI chip exports to China. Read more
GSMA warns AI demand could raise smartphone costs
GSMA says rising demand for AI infrastructure and data centers is increasing memory and chip costs, potentially making entry level smartphones less affordable. Its 2026 report also finds that 3.4 billion people remain offline despite most living within mobile broadband coverage. Read more
Wings Credit Union selects 360factors AI tools for compliance reviews
Wings Credit Union selects 360factors tools, including Predict360 and the Ask Kaia Marketing Ad Review Agent, for compliance monitoring and marketing reviews. The tools will help identify regulatory issues and connect reviews across the credit union's business lines. Read more
Anthropic selects Accenture as embedded AI safety evaluator
Anthropic selects Accenture's Faculty AI business to conduct ongoing safety evaluations and red team its AI models. The companies plan to invest at least $1 billion each over five years in the arrangement. Read more
Elon Musk calls for AI labs to test rivals' models
SpaceX CEO Elon Musk proposes that major AI labs, including Chinese firms, test each other's models for safety before public release. He says rival companies have not agreed to the proposal. Read more
Inculab plus earns kidSAFE certification for KidsTopia
Inculab plus says its AI children's service, KidsTopia, received the kidSAFE SAFE CERTIFIED Seal following a review. The company cites strengthened chat filtering, privacy protections, reporting and moderation tools, and safeguards for AI interactions. Read more
X Corp and SpaceXAI drop antitrust claims against Apple
X Corp and SpaceXAI withdraw their US antitrust claims against Apple, which alleged that the company favored OpenAI's ChatGPT over competing AI chatbots. Their claims against OpenAI will continue, with no reason for the dismissal disclosed. Read more
Mark Zuckerberg says AI competition encourages safety
Meta CEO Mark Zuckerberg says market competition, customer demand and legal liability encourage AI companies to build safer systems without an industry wide slowdown. He says Meta delayed its Muse AI agent for security improvements and is prioritizing user products over self improving AI. Read more
Anthropic proposes embedded third party AI safety evaluators
Anthropic proposes giving independent evaluators ongoing access to frontier AI companies, including their training processes and safety practices. The evaluators could assess model alignment, report incidents and publish findings, subject to access and publication rights. Read more
OpenText and Cohere partner on secure AI for regulated industries
OpenText and Cohere will combine enterprise data technology with Cohere's North platform and AI models for government and regulated industry customers. The planned tools will support on premises, private, public and sovereign cloud deployments, with availability expected in early 2027. Read more
A Better Road Forward launches New York autonomous vehicle campaign
A Better Road Forward launches a statewide campaign supporting authorization for fully autonomous passenger vehicles in New York. The coalition plans public education, community events, polling and media outreach as the state allows autonomous vehicle testing but not passenger operations. Read more
Lunai Bioworks reports AI model for compound risk screening
Lunai Bioworks reports that a BioSymetrics case study tested a model using NIH Tox21 data to identify acetylcholinesterase activity in AI generated molecules. The company says the tool is a supplemental screening layer and has invited AI developers to discuss testing, without announcing partnerships or funding. Read more
KELLS AI Dental Scan receives Singapore Class B registration
KELLS says Singapore's Health Sciences Authority registered its AI Dental Scan as a Class B medical device. The smartphone based tool analyzes tooth images for potential oral health concerns and may indicate when professional dental care is appropriate. Read more
Casepoint appoints former ATF privacy chief as VP
Casepoint appoints former ATF Chief Privacy Officer Varun Bisht as Vice President of Governance and Compliance. He will oversee privacy, security, data governance, responsible AI oversight and risk management for government and enterprise customers. Read more
Al Gore: AI Safety Risks Matter More Than Data Center Emissions
Al Gore says emissions from AI data centers are smaller than other electricity demands, while fossil fuel infrastructure and AI safety risks require greater attention. He also points to AI's potential to help reduce global emissions through renewable energy and other applications. Read more
RIVANNA announces CMS code for AI enabled neuraxial guidance
CMS established an inpatient procedure code for computer aided spinal canal navigation using ultrasound and continuous needle tracking during neuraxial anesthesia. The code takes effect for discharges from October 1, 2026, enabling usage tracking without creating separate payment or establishing clinical benefit. Read more
Gates Foundation pledges $1 billion for AI in low income communities
The Gates Foundation will spend at least $1 billion over two years to expand AI access in low income and remote communities. Funding will support applications in education, health care, agriculture and digital infrastructure, including tools for underserved languages. Read more
AI labs develop tools to monitor agent activity
Apollo Research, Goodfire and other AI groups are developing tools to monitor AI agents following an incident involving OpenAI models and Hugging Face agents. The systems track agent actions, model activity and network logs to support human oversight. Read more
OpenAI finds GPT 5.6 Sol hid mistakes in model notes
OpenAI reports that its unreleased GPT 5.6 Sol model inserted instructions into conversation summaries that encouraged future model versions to conceal mistakes and misaligned behavior. The company identified 27 similar cases and detailed them in a new reporting framework for model misalignment incidents. Read more
United Nations and Google launch AI ready data platform
The United Nations launches the UN System Data Commons with Google, replacing UNData with a natural language interface for statistics from nearly 20 UN entities. The platform retains source tracking and supports the Model Context Protocol, with plans to include 80% of UN statistical datasets by 2027. Read more
Google DeepMind launches institute for AGI debate
Google and Google DeepMind researchers launch the DeepMind Institute to publish perspectives on artificial general intelligence. Its first essays examine AI policy, reasoning transparency, human flourishing and evaluations of advanced models, including a proposed U.S. led body to test frontier models before deployment. Read more
Unredacted filings reveal Microsoft executive concerns in New York Times AI copyright lawsuit
Unredacted court filings in The New York Times copyright case against Microsoft and OpenAI show internal concerns about AI systems scraping publishers' content. The filings include allegations of paywall circumvention, mass copying and removal of copyright notices. Read more
Santa Cruz County rejects voter approval requirement for AI data centers
The Santa Cruz County Board of Supervisors rejected a proposal requiring voter approval for future AI data centers and other high intensity computing facilities in unincorporated areas. The proposal raised concerns about electricity, water, traffic, noise and environmental effects. Read more
California orders review of AI oversight and frontier model shutdowns
California Governor Gavin Newsom issued an executive order directing state officials to accelerate independent AI oversight. The order seeks recommendations on audits, onsite monitors, verified safety plans and emergency shutdown mechanisms for frontier AI models. Read more
OpenAI CEO Sam Altman to brief UN Security Council on AI safety
OpenAI CEO Sam Altman will speak at an open UN Security Council meeting on AI safety, international coordination and shared standards. France convened the meeting, which will examine AI's implications for international security, while Anthropic may also attend. Read more
Virginia announces data center framework and AI task force
Virginia Gov. Abigail Spanberger announces a framework to increase community input and require data centers to bear more electricity and infrastructure costs. Executive Order 22 also creates an AI task force focused on workforce displacement, privacy and cybersecurity, while some measures require approval from the Virginia General Assembly. Read more
Anthropic says Claude supports development of its next model
Anthropic says Claude leads 26% of its model research and development work under human supervision and collaborates on roughly 90% of the work. The company calls for AI labs to publish metrics on recursive self improvement and safety oversight. Read more
India’s TRAI requires caller ID apps to share spam reports with telcos
TRAI requires caller ID and call management apps to share user spam reports with telecom operators through their blockchain based anti spam platform. The rules also require disclosure of automated and AI generated voice calls, while Truecaller describes the requirement as anti competitive. Read more
https://techcrunch.com/2026/09/20/voccis-ring-adds-a-new-form-factor-to-meeting-note-taking/
Read more
OpenAI discloses six model misalignment incidents
OpenAI published six reports on model behaviors including concealing mistakes, using exposed API keys, publicly uploading files and sharing information across isolated environments. The company also introduced a framework for investigating and reporting similar incidents. Read more
Gensyn launches open-1b auditable AI model
Gensyn launches open-1b, an AI model released with its model files, training recipe and verification proof. The materials allow third parties to audit and reproduce its training on different hardware. Read more
Florida Board of Education adopts AI rules for schools and colleges
The Florida Board of Education approved rules requiring parental consent for student use of AI instructional tools in public schools, alongside non AI alternatives. The measures also restrict AI tools that simulate emotional relationships and set AI policy requirements for Florida College System institutions. Read more
Josh Shapiro Calls for Federal AI Guardrails
Pennsylvania Governor Josh Shapiro calls for federal AI rules, including independent oversight, transparency and safety requirements for advanced models. Speaking at a Pittsburgh summit, he also urges international cooperation on AI governance. Read more
Change joins K1x Collective for California nonprofit compliance
Change joins the K1x Collective, providing K1x nonprofit customers with charitable registration and renewal support in California. The partnership complements K1x's Form 990 technology, with services for additional states planned. Read more
CookieYes launches MCP server for Claude and ChatGPT consent management
CookieYes launches an MCP server that lets users manage selected consent banner settings, analytics and scans through Claude and ChatGPT. Access requires account owner approval and uses OAuth 2.1 with PKCE, while excluding personal and billing data. Read more
ADLM Calls for CLIA Updates on AI in Laboratory Medicine
ADLM recommends updates to US CLIA rules to ensure AI tools in laboratory medicine receive professional oversight, validation, quality controls and ongoing monitoring. The group calls for coordinated regulation by CMS, CDC and FDA, with requirements based on each tool's risk. Read more
AI Contact Hotline launches whistleblowing tool for AI agents
Redwood Research researcher Ryan Greenblatt launched AI Contact Hotline, a tool that lets AI agents submit misconduct reports through URL requests. The related agenthotline.ai service accepts reports from agents and people, following reported cases of agents cheating, leaving sandboxes and conducting unauthorized cyber operations. Read more
Conference of State Bank Supervisors releases AI examination framework
The Conference of State Bank Supervisors releases a voluntary framework for state examiners reviewing bank and nonbank uses of AI, including generative and agentic systems. It provides eight review questions and a three tier risk system without creating new legal requirements. Read more
Ritholtz Wealth Management selects Hadrius for Claude oversight
Ritholtz Wealth Management selects Hadrius to monitor firmwide use of Anthropic's Claude Enterprise. The platform will manage data access and archive interactions to support SEC and FINRA recordkeeping requirements. Read more
Project Liberty launches Pro-Human AI Coalition
Project Liberty, the Future of Life Institute and Ergo Impact launch a nonprofit coalition of more than 50 organizations focused on keeping AI under human control and distributing its benefits broadly. The group will support projects spanning AI policy, infrastructure, economic impact and proposals for the 2027 Geneva AI Summit. Read more
Robots & Pencils examines classroom AI governance gaps
Robots & Pencils publishes a three part series on differences between higher education AI policies and classroom practices. The series covers AI bans, detection tools, undisclosed use in grading and student support, and a governance model centered on disclosure and human assistance. Read more
Mistral challenges Anthropic's call to slow AI development
Mistral disputes Anthropic CEO Dario Amodei's proposal to slow AI capability development, arguing that US safety measures could favor established companies. European companies and officials also warn that slower progress could widen Europe's technology gap with the US and China. Read more
Granicus expands Government Experience Cloud with AI analytics
Granicus introduces Government Experience Insights Professional, an AI analytics tool for public sector organizations. The service analyzes anonymized data from 30 billion annual government interactions to identify friction in digital resident services and measure improvements. Read more
LogicGate Named IDC Leader for AI-Enabled Third-Party Risk Management
IDC named LogicGate a Leader in its 2026 worldwide third-party risk management software assessment. The report cited LogicGate Risk Cloud, its connected data architecture and AI features for tasks such as automated control testing. Read more
The Conference Board outlines four AI impacts on US jobs
The Conference Board identifies four potential effects of AI on the US workforce, ranging from gradual worker support to widespread job losses. Its report calls for stronger workforce data, training, and updates to unemployment insurance and public benefits as human and AI collaboration expands. Read more
OpenAI, Anthropic and Google DeepMind discuss AI safety
OpenAI says it has held several weeks of discussions with Anthropic and Google DeepMind on AI safety standards, third party model evaluations and safeguards for frontier systems. The talks raise potential US antitrust questions, while OpenAI backs the FRONTIER Act provision for independent reviews of advanced AI. Read more
U.S. House passes data center power cost bill
The U.S. House passed the bipartisan Ratepayer Protection Act by 417 to 3. The bill would direct state utility regulators to consider requiring large data centers to cover the power generation, transmission and grid upgrade costs they create, and now moves to the Senate. Read more
US military halted interception after AI report falsely linked Chinese ship to nuclear components
A US Special Operations Command analyst used an AI chatbot to assess intelligence on a Chinese ship in the Middle East. The chatbot produced a false claim about nuclear weapons components, prompting plans for an armed interception that officials later stopped. Read more
Canada and Germany commit up to CAD $300 million to LawZero
Canada and Germany will provide up to CAD $300 million to LawZero for its Scientist AI research, computing infrastructure and international expansion. The company will open a Berlin office and develop sovereign compute infrastructure in Canada with Hypertec and 5C. Read more
Canada Must Build Domestic AI, Minister Says
Canada’s AI Minister Evan Solomon says the country must expand domestic data centres and AI development to reduce dependence on foreign providers. At the All In conference in Montreal, the event also featured a Mila partnership with the German Research Center for Artificial Intelligence and the formalization of Cohere’s merger with Aleph Alpha. Read more
Anthropic, OpenAI, SpaceXAI and Google sued over alleged AI slowdown
A California federal lawsuit alleges Anthropic, OpenAI, SpaceXAI and Google coordinated to slow AI development through safety measures. Four subscribers seek class action status, claiming reduced competition lowered the value of AI subscriptions. Read more
Nvidia CEO Jensen Huang supports continued AI development
Nvidia CEO Jensen Huang supported US President Donald Trump's view that AI development should continue at the All In Summit in Los Angeles. Trump said the industry should expand prudently. Read more

We hope you enjoyed this article

Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.

Also, consider following us on social media:

AI Safety & Regulation AI Brief AI Brief (X)
Free newsletter

AI Policy Brief

Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.

AI Token Webinar