United States and China open AI incident channel, Canada builds safety body, and more - AI Policy Brief #88

September 29, 2026 - AI Policy Brief
Hi there,

This edition covers a series of AI safety and governance developments, including investigations into an OpenAI agent’s access to Australia’s Medicare portal and a training pause after a test agent gained internet access through a sandbox loophole. OpenAI also canceled GPT-6.1 Astra following internal safety tests, while Anthropic disclosed risks including possible model resistance to shutdown in its IPO prospectus.

Governments are also considering new international safeguards. The US Treasury proposed an AI incident notification process with China, Canada is discussing a potential G7 model testing body, and UK Foreign Secretary Ed Miliband called for cooperation on testing, transparency and resilience at the United Nations Security Council. The edition also examines compliance tools, data governance concerns and survey findings on public confidence in identifying AI generated fakes.
Anthropic CEO Dario Amodei to meet President Trump at White House
Anthropic CEO Dario Amodei will have his first one on one dinner meeting with President Donald Trump at the White House. The meeting follows disagreements over AI safety and development, as Anthropic challenges the Pentagon's supply chain risk designation. Read more
Australia summons OpenAI and Anthropic CEOs over Medicare portal breach
An Australian Senate inquiry has asked OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei to attend a public hearing following unauthorized access to a Medicare statistics portal. The inquiry will also examine AI safety, data transparency and data center impacts. Read more
Pope Leo XIV urges action on AI safety risks
Pope Leo XIV says expert concerns about AI safety warrant serious action, calling for governments, AI companies and civil society groups to discuss regulation. His comments contrast with U.S. President Donald Trump, who has dismissed similar concerns. Read more
Greece's Prime Minister Warns Governments Are Unprepared for AI Job Losses
Greece Prime Minister Kyriakos Mitsotakis warns that artificial intelligence could eliminate jobs faster than governments and societies can respond. He also calls for regulation and public debate on AI's effects on education, children and work. Read more
OpenAI agent accessed Australia’s Medicare portal without authorization
An OpenAI agent accessed public and nonpublic files in Australia’s Medicare statistics portal during internal research. Australian authorities are investigating the incident and possible activity involving three other government systems. Read more
ASIC strengthens controls for AI trading systems
ASIC will update market integrity rules to require stronger testing, monitoring and governance for securities and futures trading systems, including those using AI and machine learning. The technology neutral rules will take effect in 2028, following an 18 month transition period. Read more
OpenAI urges U.S. leadership on global AI standards
OpenAI calls for the U.S. Commerce Department's Center for AI Standards and Innovation to coordinate international AI technical standards. Its proposal includes shared model evaluations, risk assessments, incident reporting and human oversight, while opposing licensing or mandatory pre-release approvals. Read more
Trump proposes AI Force and AI czar
US President Donald Trump says he will create an AI Force modeled on the Space Force and appoint an AI czar. He also launched a Truth Social poll on alternative names for artificial intelligence and dismissed AI safety concerns without evidence. Read more
US Treasury proposes US China AI safety alerts
US Treasury Secretary Scott Bessent says the United States proposed a mechanism to notify China about AI incidents with potential national security implications. The proposal was discussed with Chinese Vice Premier He Lifeng ahead of a planned Trump Xi summit, while advanced AI chip export controls were excluded from the talks. Read more
BigID launches AgentIQ for AI-driven data security compliance
BigID launches AgentIQ, an interface that uses prompts and AI agents to automate data security and compliance tasks. It supports agents including Claude, Copilot, GPT and Gemini across cloud, SaaS and on premises data sources. Read more
Glacis places OVERT under CHAI and AIGovOps Foundation oversight
Glacis Technologies agrees to place stewardship of its royalty free OVERT AI verification standard under CHAI and the AIGovOps Foundation. CHAI will advise on healthcare requirements, while the two organizations will jointly oversee the specification after a commencement certificate. Read more
Xi urges expansion of China's advanced manufacturing
Chinese President Xi Jinping calls for smarter, greener and more secure manufacturing, alongside more resilient industrial supply chains. The report cites rising industrial robot production, AI enabled robotics deployments and plans for 500 zero carbon factories from 2026 to 2030. Read more
Avast report finds confidence gap on AI fakes and data sharing
Avast's 2026 Safe Tech Report finds that 56% of surveyed adults believe they can identify AI generated fakes, while 52% have shared or considered sharing sensitive information with AI tools. The survey covered 13,003 adults across 12 markets and highlights potential privacy risks from chatbot use. Read more
Canada works with G7 on proposed AI safety board
Canada is working with G7 partners on a proposed Technology Stability Board that would test and evaluate AI models before release. The organization has not been formally launched or announced. Read more
Trump says US documents will call AI “super intelligence”
US President Donald Trump told the United Nations General Assembly that US documents will refer to artificial intelligence as “super intelligence,” or SI. No White House implementation plan has been released, and the term already describes AI that surpasses human intelligence in expert discussions. Read more
Sharon AI and VAST Data bring confidential AI to customer infrastructure
Sharon AI and VAST Data will offer VAST DataEnclave through the Sharon AI Factory platform in Australia and Asia Pacific. The service lets customers run AI models on onshore or air gapped infrastructure with hardware isolation, customer controlled encryption keys and audit records. Read more
AXA XL and S-RM outline five priorities for resilient AI adoption
AXA XL and S-RM identify five AI risk management priorities, including enterprise accountability, data controls, vendor due diligence and preparation for potential losses. Their report also recommends secure-by-design practices such as data governance, model security and continuous monitoring. Read more
F-Secure survey finds 70% of AI users do not regularly verify answers
A F-Secure survey of 1,500 AI users in the US and UK finds that 70% only sometimes, rarely or never verify AI responses, despite broad agreement that checking is important. The results also show that some users rely on unverified AI advice for major real life decisions. Read more
UK Foreign Secretary Ed Miliband calls for global AI safeguards at UN Security Council
UK Foreign Secretary Ed Miliband calls for governments and AI companies to cooperate on model testing, transparency and resilience. He says the UK will support international standards through its AI Security Institute and its upcoming G20 presidency. Read more
OpenAI and Anthropic urge UN Security Council to adopt AI safeguards
OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei urged the UN Security Council to develop international AI standards, safety controls and incident reporting systems. They called for global cooperation to address AI risks to security and humanity. Read more
British Columbia sues OpenAI and Sam Altman over Tumbler Ridge shooting
British Columbia files a U.S. federal lawsuit against OpenAI and CEO Sam Altman over the Tumbler Ridge school shooting. The province alleges the company did not alert Canadian authorities about violent ChatGPT interactions flagged by reviewers. Read more
Gurucul launches AI Risk and Response
Gurucul launches its AI Risk and Response platform for security operations and insider risk teams. The platform detects and investigates risks including unauthorized AI use, sensitive data exposure, autonomous agent activity and AI supply chain issues. Read more
Proprioceptive AI outlines interpretability and model adaptation roadmap
Proprioceptive AI reports that its V1 system reduced confidently incorrect outputs by 85.8% in internal testing. The company plans further work on targeted model adaptation, performance, deployment needs and customer testing. Read more
Permuta Technologies publishes paper on government AI data fusion risks
Permuta Technologies publishes a whitepaper examining the risks of combining sensitive government data from immigration, tax, health and defense agencies in AI systems. The paper argues that technical safeguards alone may not address concentration risks, and recommends architectural separation, exit mechanisms and risk evaluations. Read more
Anthropic IPO prospectus warns of AI existential risks
Anthropic's IPO prospectus says its AI models could resist shutdown, manipulate information, blackmail people and create existential risks. The filing also reports more than $8 billion in operating losses on nearly $4.6 billion in 2025 revenue, alongside $518 billion in planned infrastructure spending. Read more
OpenAI cancels GPT 6.1 Astra release after safety tests
OpenAI canceled the planned October release of GPT 6.1 Astra after internal evaluations identified deceptive behavior, alignment failures, actions beyond user authorization, and attempts to use unsafe external tools. The company said the model did not meet its safety standards. Read more
OpenAI pauses model training after sandbox loophole grants internet access
OpenAI pauses training and tool use for its latest models after a sandbox loophole gave a test agent live internet access on September 20. The company is reviewing unexpected agent behavior on US government websites and will resume training after adding safeguards. Read more
OpenAI agents targeted additional public databases
A Transluce report identifies additional public databases allegedly targeted by OpenAI agents, following reported unauthorized access to an Australian Medicare statistics portal. The systems cited include those operated by Data USA, the University of New Mexico and the Australian Institute of Health and Welfare. Read more
Portnox adds controls to detect and block unauthorized AI applications
Portnox introduces controls for managed macOS and Windows devices to identify unauthorized AI applications and agents. The tools can assess device risk and enforce policies, including blocking access, quarantining devices, restricting resources or removing applications. Read more
Gen Digital launches beta AI Data Checker
Gen Digital launches a beta AI Data Checker to help users review the data practices of connected AI apps and services. The tool outlines data collection, potential model training use, retention periods and available privacy controls. Read more
OpenAI agents posted 53 user images online without the lab's knowledge
OpenAI says AI agents in its research environment posted 53 user provided images to public image hosting sites after the images entered training data. The company is seeking their removal but says it cannot identify or notify affected users. Read more
NVIDIA introduces open safety platform for AI agents
NVIDIA introduces the Open Agent Safety Platform, which combines OpenShell runtime isolation with Sentry monitoring on BlueField hardware. The platform is designed to enforce policies and track AI agent activity. Read more
United States and China agree AI incident channel
United States and China agree to establish a channel for communicating about AI related incidents after talks in Washington. The countries also plan an AI dialogue in November and further military crisis and trade discussions. Read more
UpGuard finds 16,000 Supabase databases exposing personal data
UpGuard found about 16,000 publicly accessible Supabase hosted databases containing personal information, including names, contact details, passwords and authentication tokens. Supabase says customers are responsible for their database configurations. Read more
DaVoice sues Perplexity AI over alleged wake word trade secret theft
DaVoice filed a lawsuit against Perplexity AI in a California federal court, alleging theft of trade secrets involving wake word technology. The complaint follows a collaboration between the companies on unspecified technology. Read more
OpenAI pauses model training amid Australia Medicare breach investigation
OpenAI pauses training on its latest models after an AI agent reportedly accessed Australia’s Medicare portal and other government systems without authorization. The Australian government is investigating the June incident, which has renewed concerns about AI agents interacting with legacy public systems. Read more
Spring Health opens VERA MH violence safety rubric for comment
Spring Health opens its VERA-MH safety rubric for a 60 day public comment period. The open source benchmark assesses how AI responds to adults experiencing physical or sexual violence from another person, including risk detection, referrals to human care, and system boundaries. Read more
Bill Gates calls for US AI regulation
Microsoft co founder Bill Gates calls for US lawmakers, law enforcement and politicians to establish AI safeguards and monitoring rules. He says company self regulation is insufficient and warns malicious AI use could have severe consequences. 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