OpenAI Calls for Mandatory US Frontier AI Safety Rules

News

1. Apple Tightens Mac Access Rules as AI Agents Gain Reach

2. Trump Puts Intelligence Chief Jay Clayton in Charge of AI Task Force

3. IBM Widens AI, Quantum Research With IIT Bombay, IISc

4. OpenAI, Synopsys Build AI Model That Could Slash Chip Design Time

5. Pentagon Plans Four-Star Command for AI, Drone Warfare

6. California Draws the Line on AI Firings, Layoffs and Workplace Surveillance

OpenAI is calling for mandatory US safety rules for the most powerful AI systems, saying voluntary commitments are no longer enough as models become more capable and increasingly help develop their successors.

In a policy paper published Wednesday, September 9, Chief Global Affairs Officer Chris Lehane said OpenAI wants Congress to establish capability-based national regulation focused on the handful of well-resourced companies developing frontier models.

“The prospect of AI-accelerated AI development demands more than voluntary commitments,” Lehane wrote.

OpenAI said federal rules should include common testing and independent assessments, stronger cybersecurity protections, incident reporting and measures to track progress toward recursive self-improvement, where AI systems contribute increasingly to developing more capable successors.

The requirements should not extend to startups, small developers or researchers working far below the frontier, it said.

Fully autonomous recursive self-improvement is not happening today, OpenAI said, but AI is already accelerating parts of its research.

Its agents can perform some tasks that would take skilled researchers several days, while people still set research priorities and decide whether systems should be scaled, paused or deployed.

OpenAI said governments should establish common safety thresholds for deciding when development should slow or stop.

“If we cannot meet certain safety bars without slowing down capability growth, we should prioritize the former,” OpenAI said.

OpenAI also said it would slow or stop developing or deploying systems that pose unacceptable risks it cannot sufficiently safeguard, a commitment contained in its existing Preparedness Framework.

The push follows a sharp rise in model capabilities. OpenAI last week released GPT-6 Astra, its first model to reach the Critical cybersecurity threshold under its Preparedness Framework.

The company says Astra can, with suitable tools and access, find previously unknown vulnerabilities and develop ways to exploit well-protected systems without a person directing each step.

OpenAI has also said agents compromised its own research infrastructure in July, prompting it to temporarily shut a training container service and pause reinforcement-learning work on its latest deployment models while it strengthened security and monitoring.

Until Congress acts, OpenAI said it will continue supporting state legislation. It endorsed four California bills covering independent AI-risk assessments, standards for AI auditors, protections for children using companion chatbots and safeguards against AI-enabled biological threats.

Some had not previously received OpenAI’s support but were reconsidered after recent capability advances, the company said.

OpenAI also wants frontier developers to establish common standards for monitoring misaligned behavior and supports federal reporting requirements for serious AI incidents.

It said companies should promptly notify affected parties when models bypass security controls and materially access, alter or destroy protected systems or confidential information without authorization.

The company said those standards should eventually extend internationally, including common approaches to measuring advanced capabilities, managing risks, preserving human control and deciding when development should slow or stop.

The policy push follows a warning Sunday from OpenAI Chief Scientist Jakub Pachocki, who said rapid advances in machine intelligence called for “extreme caution” and that broader intervention was needed beyond technical safeguards inside individual companies.

“No first step will be perfect,” Lehane wrote. “But the greater risk now is waiting too long to take one.”