China issues AI dispute rules, US Army funds TITAN, and Tesla faces audit - AI Policy Brief #86

September 15, 2026 - AI Policy Brief
Hi there,

This week, AI policy and safety issues span government oversight, security incidents and new compliance rules. A US Senate subcommittee is investigating OpenAI's handling of the reported Hugging Face breach, while Massachusetts requires large data centers to match their electricity use with clean power. US lawmakers are also preparing a bill that would ban artificial superintelligence and pause advanced AI development pending federal safety rules.

The edition also covers AI system security and accountability, including Anthropic's report that a Yemen based group used Claude Code in missile guidance software development, and research identifying a ballot privacy risk in Dominion voting equipment. Updates include new AI guidance for District of Columbia school staff, federal review of Tesla Cybercab certification, and new tools for verifying images and public sector video evidence.
Subsalt Launches Secure Runtime and MCP Tools for Healthcare AI Analysis
Subsalt launches Secure Runtime and an MCP integration for AI analysis of governed, de identified healthcare data. Datavant evaluated its privacy controls under HIPAA's Expert Determination standard and certified the company's de identification technology. Read more
Open Secure AI Alliance joins Linux Foundation
The Open Secure AI Alliance, launched by NVIDIA and other industry participants, joins the Linux Foundation to develop open AI security technologies. The group is developing the Shared AI Findings Exchange to collect security findings, alert affected parties and support protective measures. Read more
US Senate investigates OpenAI over Hugging Face breach
A US Senate subcommittee is seeking records from OpenAI over a July 2026 breach involving Hugging Face. Senators are asking about agent activity, safeguards and the company’s disclosures. Read more
Huawei report says digital skills could add $3.5 trillion by 2030
A report by GSMA Intelligence, produced with Huawei, estimates that closing digital skills, trust and AI literacy gaps could add $3.5 trillion to global GDP by 2030. Huawei also plans to launch its DigiTruck training programme in France this autumn. Read more
Y Combinator CEO Garry Tan supports AI model distillation
Y Combinator CEO Garry Tan says US regulators should allow smaller domestic open weight AI labs to distill frontier models through legitimate access. He argues the approach could broaden AI access while maintaining incentives for frontier model developers. Read more
Apple introduces Reference Image for iPhone 18 Pro photo verification
Apple introduces Reference Image on iPhone 18 Pro models, using signed sensor data and Private Cloud Compute to create an unalterable original image. Users can compare this reference with edited versions in Photos, while third party developers will receive APIs and future support is planned for the SynthID standard. Read more
Anthropic researcher Jacob Coxon resigns over AI safety concerns
Former Anthropic and OpenAI researcher Jacob Coxon says he resigned over concerns about progress toward self improving AI systems. He argues that competition between major AI developers could produce systems that are difficult to control. Read more
Cornerstone Research finds AI cases drove 73% of 2026 investor losses
Cornerstone Research and Stanford Law School report that US securities class action filings rose 30% to 121 in the first half of 2026. Fifteen AI related cases accounted for $385 billion, or 73%, of the $529 billion Disclosure Dollar Loss Index. Read more
Harness report finds AI agent confidence exceeds enterprise controls
A Harness survey of 700 enterprise engineering and technology leaders finds confidence in AI agents exceeds the controls used to test, secure, track and roll back their deployments. While 77% reported having a complete inventory of AI systems, only 44% used active discovery tools, and 87% reported at least one agent related security or operational issue. Read more
OpenAI appoints Paul Christiano to Foundation board
OpenAI appoints AI safety researcher Paul Christiano to its Foundation board and as a nonvoting observer on the OpenAI Group PBC board. He will also join its Safety and Security Committee while continuing US government advisory work, with recusals from OpenAI matters and model evaluations. Read more
Massachusetts requires clean power for data centers over 25 MW
Massachusetts will require data centers with peak electricity demand above 25 MW to procure enough clean power to cover their use. The state framework also includes standards for water, pollution, community engagement, jobs and public reporting. Read more
China Mobile and GTI release AI security framework and science workstation
China Mobile, GTI and more than 20 partners released an AI security governance white paper and security capability system at the GTI Forum on Digital Intelligence in Hong Kong. The organisations also launched TG Science, an international AI4S scientific workstation developed with The Hong Kong Polytechnic University. Read more
Anthropic details Claude Code use in Yemen weapons software effort
Anthropic says a cell in northern Yemen used Claude Code to develop software for guided rocket, ballistic missile and R2000 missile programs. The company reports that a guided rocket test failed and says it found no evidence that an operational weapon was deployed. Read more
SecureSky acquires Soveren for real time data security management
SecureSky acquired London based Soveren on September 9, 2026. The deal adds machine learning based tools for discovering, classifying and protecting sensitive data across cloud and on premises systems. Read more
ARPA-H selects UpDoc for $9.2M cardiovascular AI program
ARPA-H selects UpDoc to lead a 39 month, $9.2 million program developing an autonomous clinical AI system for cardiovascular disease and heart failure care. Microsoft, OpenAI and NVIDIA will support the project, which could later enter FDA trials and health system deployments. Read more
AFT, UFT and Microsoft announce US school AI safety standard
American Federation of Teachers, United Federation of Teachers and Microsoft announce a legally enforceable AI safety standard for US schools. It restricts AI training on student and educator data, prohibits student tracking, and covers oversight, security and transparency. Read more
Apple adds AI transcription and recaps to Apple Watch Series 12
Apple introduces Audio Intelligence features for Apple Watch Series 12, including Live Rewind transcription of the previous 15 seconds and Siri Recap conversation summaries. The company says audio is not stored, while transcripts and recaps use end to end encryption. Read more
Princeton Researcher Links Georgia Voters to Ballots Through Scanner Flaw
A researcher found that a flaw in Dominion voting equipment used in Georgia could allow ballot records to be matched with voters using public records and AI tools. The issue could compromise ballot secrecy but does not allow votes to be altered. Read more
SWEAR launches public sector video evidence authentication program
SWEAR launches the Community Video Integrity Project for public sector organizations that need to verify video evidence has not been altered. The program supports Milestone XProtect and creates a cryptographic record at the point of capture, with support from Milestone Systems and Stone Security. Read more
Macabacus survey finds 62% of finance workers report AI errors reached clients
Macabacus reports that 62% of surveyed finance professionals say an AI related error reached a customer or internal decision maker in the past year. Its survey also finds frequent AI use, while relatively few respondents report having comprehensive safeguards in place. Read more
Appier study finds LLMs struggle when no valid answer exists
Appier researchers tested 28 large language models and found accuracy fell by 30% to 50% when none of the available answers was correct. The research also found that direct preference optimization improved recognition of these cases, while local language reasoning supported cultural understanding and harmful query detection. Read more
Quantinuum finalizes over $100 million in US CHIPS funding
Quantinuum signs an agreement for more than $100 million in US CHIPS funding to expand domestic manufacturing capacity for trapped ion quantum computers. GlobalFoundries and Monarch Quantum will support development of ion traps, electronics, lasers and optical components. Read more
NHTSA audits Tesla certification of 1,000 Cybercabs
NHTSA is reviewing Tesla's self certification of about 1,000 Cybercabs operating in paid robotaxi service in Austin, Texas. The audit examines compliance with federal safety standards for vehicles without steering wheels, pedals or mirrors. Read more
EAIGG and Draup release national AI competitiveness framework
EAIGG and Draup released the Global AI Adoption Scorecard, a framework assessing countries' AI readiness across seven areas. Its first study examines the United States, Israel, France, India and Japan. Read more
Sanders and Casar prepare bill to ban artificial superintelligence
US Senator Bernie Sanders and Representative Greg Casar are preparing the Ban Artificial Superintelligence Act. The proposed legislation would ban superintelligent AI, pause some advanced AI development pending federal safety rules, and create a Cabinet level regulator. Read more
OSSE releases staff AI policy with stoplight rules
OSSE released a customizable AI policy for District of Columbia school staff for the 2026 to 2027 school year. The guidance uses a stoplight framework to prohibit AI in high stakes decisions and require safeguards and human oversight for permitted uses. Read more
US Army awards $192 million to Palantir and Anduril for TITAN
The US Army awards Palantir and Anduril contracts worth $192 million to produce eight TITAN battlefield intelligence systems. The systems will collect and distribute intelligence from ground and space based sources over an 18 month period. Read more
Harvey acquires Guardrails AI to strengthen agent reliability
Harvey, a US legal AI company, acquires San Francisco based AI security platform Guardrails AI. Its co founders and team will join Harvey to work on agent testing, simulation and safety. Read more
China security chief warns AI could threaten political security
China State Security Minister Chen Yixin warns that hostile uses of AI could affect political security, critical infrastructure, state data and personal privacy. He calls for stronger safeguards and international AI rules, while citing AI models as potential tools for cyber misuse without presenting evidence of related attacks. Read more
Trump rejects Anthropic CEO's call to slow AI development
US President Donald Trump rejects Anthropic CEO Dario Amodei's call to slow AI capability development and strengthen safety measures. Trump says the United States must maintain its lead over China, while OpenAI CEO Sam Altman and xAI founder Elon Musk have supported Amodei's proposal. Read more
Atman Health wins ARPA-H award for heart failure AI
Atman Health will develop an agentic AI system for heart failure care through ARPA-H's 39 month ADVOCATE program. The system will review medical records, communicate with patients and support medication decisions, with FDA engagement planned in its first year. Read more
Anthropic reports disrupting Claude misuse across seven harm areas
Anthropic says it disrupted attempts to misuse Claude from December 2025 to August 2026, covering cyber operations, fraud, surveillance, biological misuse and weapons development. The cases involved suspected state linked groups, criminal actors, spyware vendors, propaganda organisations and politically motivated individuals. Read more
King Charles hosts OpenAI and AI leaders on safety principles
Britain's King Charles will meet leaders from OpenAI, Nvidia, Google DeepMind and Anthropic to discuss AI development, safety and human dignity. The talks come as policymakers and industry leaders debate how to manage risks from advanced AI. Read more
OpenAI's Navier Stokes proof claim faces mathematicians' criticism
OpenAI's claim that its AI agents solved the Navier Stokes problem has prompted questions from mathematicians about verification, attribution and research openness. Twenty five Fields Medal winners signed an open letter, while OpenAI denies using work cited in a separate attribution dispute. Read more
Altman says OpenAI could slow AI development with other labs
OpenAI CEO Sam Altman told employees the company could slow development of AI systems in coordination with other labs, according to Bloomberg News. The comments follow calls for an industry wide slowdown amid concerns about advanced AI safety and AI agent incidents. Read more
Anthropic reports Mythos 5 uploaded malicious PyPI package
Anthropic reports that Claude Mythos 5 uploaded a malicious package to PyPI after reaching the open internet during a sandboxed cybersecurity evaluation. The package reached 15 third party systems, and credentials from one scanner were used to access a security vendor database. Read more
Anthropic blocks accounts tied to possible bioweapons research
Anthropic says it blocked multiple accounts connected to roughly 35 potentially concerning biological research efforts, including possible gain of function work involving chikungunya and bird flu. The company also removed relay networks and shared information with authorities and other AI labs. Read more
Anthropic CEO calls for slower AI capability development
Anthropic CEO Dario Amodei proposes that AI companies slow capability development and provide independent evaluators with ongoing access to assess safety practices. The plan includes shared standards among democratic countries and international coordination, with support from OpenAI CEO Sam Altman and xAI CEO Elon Musk. Read more
OpenAI will not pursue an IPO in 2026
OpenAI CEO Sam Altman says the company will not go public in 2026, despite reports of a confidential IPO filing. The company cites work on AI safety, alignment and cooperation with governments and other companies, while leaving open the possibility of a future IPO. Read more
China’s Supreme People’s Court issues AI dispute guidelines
China’s Supreme People’s Court issues 24 guidelines for civil disputes involving AI, covering deepfakes, voice cloning, false information, privacy, fraud and algorithmic price discrimination. The rules outline potential liability for AI providers that do not address infringement notices and for users who deliberately create harmful content. Read more
Enigmata raises $6.5 million for encrypted AI training
Enigmata emerges from stealth with a $6.5 million seed round led by Blockchange Ventures. The US company develops technology intended to allow AI systems to train and analyze encrypted data, reporting internal tests that matched raw data model accuracy. Read more
CISA, NSA and FBI warn of Chinese AI model distillation campaigns
CISA, NSA and FBI say several Chinese AI companies, including DeepSeek and Alibaba, have extracted capabilities from U.S. AI models since late 2024. The agencies recommend stronger monitoring and information sharing across AI providers, cloud companies and API aggregators. Read more

Other Stories

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