AI Governance & Commercialization: OpenAI Expands Policy Scope While Rivals Race to Integrate Agents

Policy & Safety

• OpenAI is strengthening its leadership team in preparation for an IPO by hiring Noam Shazeer and Dean Ball. Shazeer, a co-lead at Gemini and founder of Character AI, will join OpenAI after announcing his departure from Google DeepMind AI on Wednesday [5]. Furthermore, Ball, who previously helped publish America’s AI Action Plan during a brief stint in the White House, will lead OpenAI's new Strategic Futures team to shape frontier AI policy [5]. This team will report directly to Chief Strategy Officer Jason Kwon and focus on matters including catastrophic risk, recursive self-improvement, and the relationship between the company, governments (especially the U.S. Federal Government), and society [5]. ^[5]
• Following alleged ties between SK Telecom, a company with major business interests in China, and US national security concerns, Anthropic was forced by the White House to cut off access to its AI models Claude Mythos and Fable 5 for all foreign nationals, including users inside the US and its own employees. The initial alarm stemmed from SK Telecom's alleged ties to China, which prompted the company to comply with the directive [8]. This unprecedented use of export controls to restrict access to an entire AI model—rather than requiring a physical transfer of source code or weights—has exposed significant instability in current US AI governance and has led experts to question the sustainability of existing regulatory regimes [43]. ^[43][8]
• Google Deepmind is developing an advanced 'AI Control Roadmap' that treats its internal AI agents as potential insider threats, building a safety framework modeled after cybersecurity tools like MITRE ATT&CK [11]. This system uses trusted AI supervisors to monitor the reasoning and actions of active agents, with prevention systems designed to block high-risk actions before they occur [11]. The roadmap details four escalation levels for monitoring (D1 through D4) and three response levels (R1 through R3), mapping out defenses as AI capabilities increase and the risk of evasion grows [11]. ^[11]