OpenAI gives board veto power over risky AI

OpenAI is trying to prove it still cares about safety after last month’s chaotic boardroom brawl.

OpenAI released an updated "Preparedness Framework," detailing how it plans to track, evaluate, and prevent catastrophic risks from cutting-edge AI models. The biggest takeaway? The board of directors now officially has the power to veto risky model releases—even if CEO Sam Altman wants to ship them.

Defining catastrophic risk

OpenAI defines catastrophic risk as anything that could cause hundreds of billions of dollars in economic damage, or lead to the severe harm or death of many people. This covers everything from sci-fi existential threats down to real-world chaos.

The framework tracks four primary risk categories for frontier models: cybersecurity, persuasion (like coordinated disinformation), model autonomy (agents acting outside human control), and CBRN (chemical, biological, radiological, and nuclear threats).

Models are graded on risk levels after accounting for safety guardrails. If a model lands in the "high" risk tier, it cannot be deployed to the public. If it hits the "critical" tier, OpenAI pledges to stop further development entirely.

Take cybersecurity: an AI that merely boosts the productivity of security analysts is considered medium risk. If it can autonomously develop working exploits against hardened targets without human input, that is high risk. If it can devise and execute an end-to-end cyberattack from scratch given only a vague prompt, it is deemed critical and shelved.

A new safety group and board vetoes

To keep engineers from grading their own homework, OpenAI is establishing a cross-functional Safety Advisory Group. This team will review technical assessments and issue recommendations to leadership.

Crucially, those recommendations must be sent simultaneously to executive leadership (including Sam Altman and CTO Mira Murati) and the board of directors. Leadership decides whether to ship or freeze a model, but the board retains the ultimate authority to overturn that decision.

This setup looks directly tailored to prevent the rumored tensions that sparked November’s boardroom revolt, where directors reportedly felt kept in the dark about high-risk projects. However, the dynamics have fundamentally shifted: vocal safety advocates like Helen Toner are gone, Ilya Sutskever’s role is altered, and business-friendly figures like Larry Summers and Bret Taylor now hold the board seats. Whether this board will actually pull the emergency brake remains to be seen.

Why it matters

For businesses and developers building on OpenAI, this framework offers a clearer operational boundary. You won't have to worry as much about sudden, catastrophic liabilities emerging from API upgrades, though it also means certain high-autonomy features will be deliberately withheld.

For the AI landscape, OpenAI is setting a precedent by codifying explicit technical thresholds for when a model must be shut down. Expect competitors like Anthropic and Google to face intense pressure to match this level of documented commitment, giving regulators an off-the-shelf blueprint for AI governance.

We think this move is a politically savvy defense mechanism. It rebuilds public trust by showing structured oversight, while conveniently reinforcing OpenAI's core narrative that its models are almost dangerously powerful. The clear vulnerability, however, is transparency: outside of vague promises for third-party audits, the entire process remains behind closed doors. If a model gets killed or a veto gets overridden, the public might never know.