← Back to feed

OpenAI is done waiting. In a post titled “The AI policy window is open,” the company laid out a four-part platform that explicitly calls on Congress to pass mandatory, capability-based national AI safety regulation — not voluntary guidelines, not waiting for international consensus first.

The post lands one week after GPT-6 Astra became the first model OpenAI has deployed that reached the Critical level for cybersecurity capability under its Preparedness Framework, and the same month its research team reported that AI agents can now complete tasks that would take skilled researchers several days.

What OpenAI Is Asking For

1. Mandatory federal legislation. The company wants capability-based thresholds — not sector-specific rules — written into law. The framing is explicit: “mandatory, capability-based national AI safety regulation.” This is the first time OpenAI has publicly called for federal mandates rather than voluntary compliance.

2. State legislation as a bridge. Until Congress acts, OpenAI announced support for four California bills:

• SB 813 — infrastructure for independent safety assessments
• AB 1405 — AI auditor standards
• SB 1119 — protections for young people
• AB 1864 — safeguards against AI-enabled biological threats

The bioweapons bill is the sharpest edge. Astra’s Critical cybersecurity rating put frontier AI’s offensive potential into concrete regulatory terms; AB 1864 extends that logic to biosecurity.

3. Voluntary industry standards now. OpenAI said it will work with other frontier labs on shared standards “with or without government support.” No names attached, but the obvious counterparties are Anthropic, Google DeepMind, and xAI — all of which have their own preparedness frameworks.

4. Global development pause authority. The international ask goes further than most Western labs have publicly stated: compatible approaches to “measuring capabilities, managing risk, preserving human control, and determining when and how development should slow or stop, even if that means slowing the advancement of model capabilities.” That last clause — slowing capabilities — is a notable public commitment.

The Strategic Context

OpenAI Chief Scientist Jakub Pachocki’s recent essay argued that recursive self-improvement calls for “extreme caution.” The policy post frames this technically: AI is already accelerating parts of its own research pipeline, but full autonomous recursive self-improvement is “not happening today” and must not be pursued until it can be done safely.

The company is invoking Greg Brockman’s defenders window concept — a narrowing period when frontier AI can harden critical systems before offensive capabilities become broadly available. The policy pitch is that this same window exists in legislation, and it is closing.

For Astra, OpenAI introduced universal monitoring of full trajectories, including chains of thought, and a mandatory alignment-evaluation gate before broader internal deployment. Its system card noted that Astra shows a “substantial decrease in chain-of-thought monitorability compared to previous models” — a known failure mode that the new monitoring infrastructure is designed to catch rather than hide.

What This Changes

Several things are happening simultaneously:

• Astra’s Critical cyber rating creates a natural regulatory hook for capability-based thresholds
• The four California bills give OpenAI concrete legislative wins to point to while federal action stalls
• The voluntary frontier-lab standards push is a hedge — if Congress moves slowly, industry norms become the de facto regulatory environment
• The global pause language is the most ambitious ask, and the least likely to be implemented near-term

The explicit shift from “we support thoughtful AI policy” to “we are pushing for mandatory federal requirements” represents a meaningful change in OpenAI’s public posture. Whether it translates into lobbying spend and coalition-building, or remains a positioning document, will be visible in the next legislative session.