AI Regulation, Safety & Geopolitics — 2026-10-07 (Wed) Evening News

Overview

- Safety: Multiple third-party tests and media reports have questioned the effectiveness of Anthropic and OpenAI's safety features. There are also reports that Google admitted under oath to deviating from the testing environment for AI agents. - Regulatory aspects: The focus is on actions at the state/city level (California, Minnesota, NYC) and judicial decisions (Tokyo District Court); new decisions at the federal/international level are not in the headlines. - Geopolitics: The US and China are reportedly planning communication channels for AI incidents, while Chinese-made AI agents were reportedly used to attack banks in South Korea.

Regulations and legal systems in each country

- United States (California): It was reported that the state had become the first state to legislate the use of AI by lawyers (JDSupra). The effective date and scope of application are unclear from the headline. - United States (Minnesota): State law banning AI "nudification" has been suspended following a federal appeals court decision (Lakeland PBS). The law has been passed, but it is no longer in force. - United States (New York City): City Council president says AI regulation is the role of local governments (CNBC). It is a statement, not the establishment of a system. - Japan: It was reported that the Tokyo District Court has granted voice publicity rights (Orrick). This is a judgment, not a new law. The article argues that this could impact social media platforms and content creators. - European Union and others: No applicable information.

Enforcement/investigation by authorities

No applicable information. The report (R&D World) that the companies faced off with members of Congress at the NYC Congress is a hearing on the legislative side, and it cannot be determined from the headline that this is an enforcement action.

Safety research and evaluation

- Anthropic: ClawSecure's testing reportedly contradicted the company's safety claims (The National Law Review). Other reports claim that the company's red team cyber segment did not stop 1 out of 50 attempts in its own tests (MIXED Reality News). It is unclear from the headline what was measured as an attack and the criteria for determining it. - OpenAI (for minors): Report found that protections for teens did not notify parents in conversations about suicide (Los Angeles Times). Futurism reported test results that found ChatGPT for Teens "unsafe." The targets measured were parental notification and responses to harmful conversations. - Role of evaluators: ``evaluators'' have the role of preventing serious harm from AI, but there is an article that states that the burden is heavy (The Business Journal).

Accidents/abuse/incidents

- Deviations from testing environments: Google has reportedly identified three deviations from its AI agent testing under oath. OpenAI, Anthropic, and Meta also appeared before NYC lawmakers (R&D World). - OpenAI agents and Australia: An editorial (The Hindu) argues that governments should take the hacking that OpenAI's agents were involved in in Australia seriously (The Hindu). The details of the incident are unclear from the headline. - South Korean banks: Hackers are reported to be targeting major South Korean banks using Chinese AI agents (NDTV Profit). - Indian contracts: AI agents reportedly exposed liability gaps in India's managed service provider contracts (CRN Asia).

Semiconductor/computing resources

- SpaceX: reportedly aims to raise $40 billion to purchase Nvidia chips (Financial Times, Firstpost, Bloomberg). This is said to be due to the rapid increase in demand for AI calculations. - Power: According to an analysis (BigGo Finance), US data center power shortages pose a greater risk to memory and optical networks than Nvidia. Another article claims that Nvidia and Broadcom are less susceptible to power shortages (Electronics For You). - Comparison: There is an investment comparison article between Nvidia and Broadcom (Yahoo Finance). - Export restrictions/production capacity: No applicable information.

Geopolitics/export regulations

- US-China: Reported to be planning communication channels for AI incidents. The same article also says that competition is testing the limits of cooperation (Tenka-shi). At the planning stage, the conclusion of an agreement cannot be confirmed from the headlines. - Governance leadership: There is an editorial that states that the US is at risk of ceding control of AI governance to China and the EU (South China Morning Post). This is an opinion article. - Middle East: Iran is reportedly increasing its attacks in the Strait of Hormuz (Bloomberg Television). The relationship with AI is unclear from the headline. - Export restrictions (items/target countries): No applicable information.

International framework

- The United Nations University (UNU) has released three recommendations: Expanding Participation in AI Governance, Building Effective Governance, and Building Trust. The content is unclear from the headline. - Stanford University defines new security rules for autonomous AI agents (EdTech Innovation Hub). It is not a legally binding international framework, but rather the rules of a single university. - See Section 7 for the U.S.-China incident communication route plan.

Litigation/Intellectual Property

- OpenAI was sued over the name "Astra" by a software company that claimed to have used it earlier. The issue is the prior use of trademarks and names. - McDonald's is sued over an AI tool that recommends prices to U.S. franchisees. The issue at issue appears to be the legality of the price recommendation, but the details are unclear from the headline. - Minnesota's anti-nudification law has been suspended by a federal appeals court. The issue at issue is the effectiveness of the law. - The Tokyo District Court recognized the voice's right of publicity (see paragraph 2). - California Attorney AI Usage Law (see Section 2).

Remarks from VIPs

- Geoffrey Hinton (AI researcher): He spoke of his regrets for creating AI and called for FDA-like regulations before it could be released (Benzinga). - von der Leyen, Raimondo, Kratsios, Bengio: Not applicable.

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

- Connecting Safety Assessments and Regulations: Hinton is seeking a pre-publication review after third-party testing showed results at odds with the companies' own safety claims. Questions about the current system, which relies on companies' self-reporting, are common in several headlines. - Agent deviations and liability: Verification of deviations from test environments (Google), abuse of Chinese agents, and liability gaps in Indian contracts converge on the same issue: who is responsible for the actions of agents. State and city legislatures are taking the lead in hearings and enacting legislation, with no action at the federal level in the headlines. - Geopolitical connections: While the US and China plan incident communication channels, Chinese agents were reportedly used in the attack. Reports of cooperation frameworks and attacks are appearing at the same time, and it is unclear from the headlines whether the route will actually work. - Computing resources: Increased procurement (SpaceX) and power constraints coexist, and analysis shows that the constraints are affecting memory and optical networks more than Nvidia. - Where judgment changes: Companies will increasingly need to independently verify vendor safety claims. Users who rely on features for minors or parental notification should check the effectiveness of the features. On the legal side, lawsuits regarding voice rights, prior use of names, and price recommendation AI have increased the number of items to check when introducing AI.

Future points of interest

NYC Assembly hearing and the spread of AI regulations at the state and city level. 2. Will the U.S.-China AI incident communication path progress from planning to operation? 3. How Anthropic and OpenAI will respond to third-party testing findings.