AI slowdown pact, OpenAI shelves IPO, and 3 more AI moves

In brief

• Frontier safety pacing: Anthropic CEO Dario Amodei proposed a three-step framework to slow frontier AI capabilities, drawing immediate public agreement from OpenAI CEO Sam Altman and xAI founder Elon Musk.
• Capital markets divergence: OpenAI ruled out a public listing in 2026, with Sam Altman stating that proceeding while catastrophic risks remain unresolved would be ill-advised.
• Global South open-source AI: Chinese President Xi Jinping proposed a BRICS AI Open Source Zone in New Delhi to coordinate large language model training and shared technology stacks across expanded member states.
• Strategic technology protection: South Korea's revised Criminal Act entered into force, establishing a mandatory minimum sentence of three years in prison for leaking AI, chip, and battery technologies to foreign entities.
• Autonomous fleet deployment: Hyundai Motor Group postponed its proprietary Atria autonomous software to 2029 and partnered with Nvidia to deploy interim Level 2+ and Level 2++ driver-assistance systems in 2028.

Five developments

1. Anthropic CEO Dario Amodei proposes three-step plan to pace frontier AI development

• Anthropic Chief Executive Dario Amodei published a policy essay on September 12 calling on artificial intelligence labs to deliberately pace capability gains, warning that self-improving agent swarms could become capable of disrupting internet infrastructure within 6 to 12 months 1.
• The proposal outlines three measures: placing embedded third-party evaluators inside frontier labs with employee-level system access, establishing coordinated industry safety standards among leading developers, and securing multilateral cooperation among democratic governments while preserving a technological lead over authoritarian states 1.
• OpenAI Chief Executive Sam Altman and xAI founder Elon Musk publicly endorsed the framework on X, with Amodei noting that domestic coordination among frontier labs will require targeted U.S. antitrust exemptions to allow safety collaboration without competitive penalties 1.

2. OpenAI rules out 2026 IPO as Sam Altman calls AI extinction risk 'unacceptable'

• OpenAI will not pursue an initial public offering in 2026 despite its confidential registration filing, Chief Executive Sam Altman said in remarks published on September 12, citing urgent safety and alignment responsibilities 2.
• Altman stated that even a single-digit to 10% risk of catastrophic outcomes from frontier AI systems is intolerable, emphasizing that commercial pressures and public-market milestones must not override containment and safety verification 2.
• The postponement creates a clear strategic divergence with rival Anthropic, which continues preparing to market a $100 billion offering valuing the company at approximately $2 trillion ahead of the November U.S. midterm elections 2.

3. Xi Jinping proposes BRICS AI Open Source Zone at New Delhi Summit

• Chinese President Xi Jinping proposed the creation of a BRICS AI Open Source Zone during the closing session of the annual leaders summit in New Delhi on September 13, urging member nations to coordinate on foundation models and supply chain stability 3.
• The initiative aims to build shared large language model training pipelines, common open-source developer ecosystems, and joint digital infrastructure across expanded BRICS members, including the UAE, Iran, Indonesia, Egypt, and Ethiopia, alongside 10 partner nations 3.
• The multilateral proposal aligns with Indian Prime Minister Narendra Modi's summit address advocating accessible, non-aligned technological solutions tailored to the Global South, creating an alternative framework to Western compute and model ecosystems 3.

4. South Korea's expanded espionage law takes effect with three-year minimum for AI and chip leaks

• South Korea's revised Criminal Act took effect on September 13 following a six-month grace period, expanding the legal definition of espionage beyond actions benefiting North Korea to encompass all foreign countries and entities 4.
• The statute creates a dedicated offense for unauthorized transfers of strategic national assets, establishing a mandatory minimum prison sentence of three years for leaking technologies across artificial intelligence, advanced semiconductors, memory architectures, and batteries 4.
• The National Intelligence Service welcomed the enactment, which closes a long-standing statutory loophole that previously forced prosecutors in corporate technology theft cases to rely on trade secret statutes carrying lower penalties 4.

5. Hyundai delays in-house autonomous software to 2029 and partners with Nvidia for 2028 deployment

• Hyundai Motor Group announced on September 13 that it will delay commercial deployment of its proprietary Atria autonomous driving software to late 2029, turning to Nvidia to supply interim Level 2+ and Level 2++ driver-assistance systems in 2028 5.
• The interim automated driving architecture will run on Nvidia's Hyperion 10 compute platform using an array of cameras, radar, and ultrasonic sensors rather than lidar, with lidar hardware reserved for future Level 3 systems 5.
• Together with affiliate Kia, the world's third-largest automaker plans to use its global fleet of over 7 million annual vehicle deliveries as a data flywheel, aiming to accumulate more real-world autonomous driving data than competitors by 2033 while refining its proprietary software stack 5.

The read-through

Watch next

• Antitrust safe-harbor reviews: Responses from the U.S. Department of Justice and Federal Trade Commission regarding antitrust clearance for frontier lab safety coordination.
• Anthropic prospectus filing: Disclosures in Anthropic's pending SEC registration regarding independent evaluator access and anchor syndicate commitments.
• BRICS working group specifications: Structural guidelines and compute-sharing protocols for the proposed AI Open Source Zone ahead of the 2027 Beijing presidency.