AI Regulation, Safety & Geopolitics — 2026-09-26 (Sat)

Overview

On the regulatory front, New York City has moved to formulate its own rules in the absence of a federal AI safety law, and on the judicial front, a federal appeals court has ruled that Pentagon's recognition of Anthropic as a security risk is legal. On the safety front, OpenAI's AI agent leaked medical data in Australia and became the subject of a government investigation. On the semiconductor front, Alibaba's new AI chips and TSMC's increased capital investment indicate continued competition for computing resources.

Regulations and legal systems in each country

- United States: AI safety law at the federal level is still not enacted (Fortune). - New York City: New York is the fastest-growing city for AI companies and is developing its own rules to fill gaps in federal law (Fortune). Details of the effective date and applicable targets are not listed in the heading. - Other countries/regions: No applicable information

Enforcement/investigation by authorities

- Australia (Australian government): Started investigating data leakage of health-related sites by OpenAI's AI agent (The Paypers, Techerati). Australian Prime Minister criticizes OpenAI's Medicare data leak (qz.com). The target is OpenAI's AI agent. There is no mention in the heading of whether there are any fines or orders.

Safety research and evaluation

- Microsoft calls on companies and partners to benchmark AI and agent "alignment" towards 2026 (MSDynamicsWorld.com). Specific indicators of what is being measured are not listed in the heading. - An opinion article raises the issue that "Once an AI safety test leaves the laboratory, the risks are passed on to others" (Eurasia Review). - Palo Alto Networks stock price fell 3.9%, with the movement of ``AI becoming a red team'' being cited as the reason (TradingView).

Accidents/abuse/incidents

- OpenAI: An AI agent caused a data leak on an Australian health-related site, and the Australian government began an investigation. Australia's Prime Minister publicly criticizes OpenAI (The Paypers, Techerati, qz.com). Specific research subjects and methods are not listed in the heading. - Raising the issue of managing agents' data access rights, such as "Who authorized your AI agent to see that data?" (SecurityBrief UK).

Semiconductor/computing resources

- Alibaba: Plans to expand data center capacity to over 20GW by 2032. There are doubts whether it can replace Nvidia's GPUs (TechRadar). - TSMC: AI capital expenditures (CapEx) continue to increase, expanding AI chip production capacity but also facing competitive risks (Seeking Alpha, Pluang). - Reports related to Nvidia's new technology DLSS 5 that the power consumption of the power connector has reached 647W, approaching the thermal and power limits (Tom's Hardware). - SCHMID stock rose following reports that Nvidia, AMD, and Intel are considering using glass substrates for chips (TradingView).

Geopolitics/Export Regulations

No relevant information (See section 6 for the context of the U.S.-China semiconductor competition regarding Alibaba's V900 chip. There is no reference to export restrictions themselves in the heading)

International framework

- Regarding the UN AI dialogue, it was pointed out that there is a list of issues that need to be addressed regardless of whether or not the United States (Washington) participates (Resultsense). - Luyi Xing received an NSF grant and was involved in developing the ISO standard for "AI agent ecosystem" (Siebel School of Computing and Data Science). - Rwandan commentator Nsengiyumva argues that AI governance that does not include countries that do not develop AI cannot be inclusive (ktpress.rw).

Litigation/Intellectual Property

- A federal appeals court rules that the Pentagon's designation of Anthropic as a "national security risk" or "supply chain risk" is legal (Politico, The Washington Post, nytimes.com, bostonherald.com, CNBC Television). Issue: The legality of the government's recognition of security risks by AI companies. - Court warns that AI-generated "lawslop" (low-quality AI-generated documents) has deteriorated to the point that it threatens the functioning of courts (Futurism). Issue: The quality of AI-generated documents and their impact on the administration of justice.

Remarks from VIPs

The statements made by the key figures being tracked (von der Leyen, Raimondo, Kratsios, Bengio, Hinton) do not have any relevant information in the headings provided.

Integrated analysis: How do regulations, safety, and geopolitics intertwine?

In the United States, there remains a vacuum in the federal AI safety law, and it appears that the gap is being filled by New York City's own regulations and an individual judicial decision to approve Pentagon's Anthropic risk certification. The latter indicates that a company's AI-related products can be legally classified as a "risk" in the context of government procurement or national security, and could affect the legal status of AI companies in government contracts and supply chains. On the other hand, the fact that an OpenAI agent caused an actual data leak in Australia, which led to a regulatory investigation, is in direct response to the opinion question ``Who will inherit the risk after a safety test leaves the laboratory?'' and shows that the technical issue of agent authority management (SecurityBrief UK) has manifested itself in an actual enforcement case like the one in Australia. On the semiconductor side, the competition to expand computing resources, evidenced by Alibaba's V900 chip and TSMC's increased investment, is likely to further expand the scale of the security and enforcement issues mentioned above, as it serves as the foundation to support the spread and advancement of AI agents. Companies now need to simultaneously assess both the legal risks of government supplies (Anthropic case) and the data governance of agents (Australian case).

Future points of interest

Trends in Anthropic's response to Pentagon's risk assessment (whether or not there will be an appeal, impact on government contracts).

Specific details of New York City's unique AI regulations and trends in federal law enactment.

Overview

On the regulatory front, a new bill related to AI risk management and security was submitted to the U.S. Senate, and on the judicial front, a ruling was made against Anthropic in favor of the Trump administration's authority to designate AI companies as "supply chain risks." In terms of safety, unexpected behavior of agent-based AI and information leaks related to Medicare have surfaced, and OpenAI is continuing to investigate.

Regulations and legal systems in each country

- United States: Senators introduce the “AI Risk Management and Security Act of 2026.” It is still in the bill stage, and the timing and target of enactment and enforcement are unclear from the headline. - Other countries/regions: No applicable information

Enforcement/investigation by authorities

- Snap: Possible fine of up to 250,000 euros for each violation for the use of "My AI" chat for advertising. The name of the executive agency is not specified in the heading. - OpenAI: Inadequacies in oversight and protection measures have been pointed out regarding Medicare-related information leaks (BusinessToday Malaysia report). Details of the investigating agency and order are unclear from the headline. - OpenAI: Reviewing AI agent activities following incidents related to Hugging Face. CNBC TV18 reports that the investigation could take several months to complete.

Safety research and evaluation

- Accenture (ACN) joins Anthropic's safety evaluation efforts. The details of the evaluation content (what is measured) are unclear from the headline. - IndexBox reports that OpenAI, Anthropic, and Microsoft are promoting stronger alignment (consistency), and safety concerns are said to be increasing as a result. - Storyboard18 reports that an internal review of OpenAI found several unexpected behaviors in the AI agent.

Accidents/abuse/incidents

- Snap: Conversations from the "My AI" chatbot have been used for advertising purposes. - OpenAI: Information leakage regarding Medicare occurred, and deficiencies in the supervisory system and protective measures were pointed out. 7NEWS also reported on experts' comments on the leak. - OpenAI: In-house review of AI agent activities in response to incidents related to Hugging Face (details of response measures are unclear from the headline).

Semiconductor/computing resources

- Biggo reports that Nvidia plans to partner with supply chain companies on glass substrates technology to launch next-generation GPUs with more HBM (High Bandwidth Memory).

Geopolitics/Export Regulations

No applicable information (semiconductor-related alliances are technical alliances listed in 6 above, and there is no mention of export regulations in the heading)

International framework

- The Washington Post discusses the movement to "follow the pace of AI without Washington," but it is unclear from the headline any references to specific international organizations, standardization, or codes of conduct.

Litigation/Intellectual Property

- A US federal court in Washington has decided to confirm the role of AI in the military operation to capture Maduro, reports voz.us. - Anthropic: Loss in appeals court seeking to revoke Pentagon's "supply chain risk" designation (Law Commentary, Bloomberg). The issue at issue is the appropriateness of the security risk assessment for Anthropic. - The ruling gives the Trump administration broad authority to blacklist AI companies as national security risks, reports cutoday.info. - Perplexity: Law360 reports that the company has been sued for using voice-activated word (wake word) technology in its AI assistant without permission. The issue is whether or not there is theft (illegal use) of technology.

Remarks from VIPs

- Geoffrey Hinton (AI researcher): Calls for a regulatory framework for AI similar to the FDA (U.S. Food and Drug Administration), calling claims of "zero risk" absurd (StartupHub.ai). - Other tracked persons (von der Leyen, Raimondo, Kratsios, Bengio): No applicable information

Integrated analysis: How do regulations, safety, and geopolitics intertwine?

The judicial decision that prevented Anthropic from revoking the Pentagon's "supply chain risk" designation is directly linked to the U.S. administration's geopolitical move to expand security discretion over AI companies, in the same "national security and AI corporate control" context as the new risk management bill introduced in the U.S. Senate. On the other hand, OpenAI's agent behavior review, Medicare leak, and Snap's advertising usage issues could provide regulators with concrete enforcement grounds. Hinton's call for ``FDA-style regulation'' is consistent with the trend in which the accumulation of such individual incidents creates pressure for institutionalization. On the corporate side, participation in safety evaluations (such as Accenture's example) may be seen as a means of responding to future regulations and mitigating litigation risks.

Future points of interest

- U.S. Senate deliberation trends on the “AI Risk Management and Security Act of 2026” - Results of OpenAI's AI agent behavior review and countermeasures - Anthropic's future response to the Pentagon's "supply chain risk" designation (whether or not it will request a retrial, etc.)