OpenAI, Anthropic, and Google are trying to agree on AI safety standards

The CEOs of OpenAI, Anthropic, and Google DeepMind have been in discussions about forming a shared industry standards body for AI model safety protocols.

Sam Altman, Dario Amodei, and Demis Hassabis have each publicly called for US-led collaboration on AI safety standards, including international forums for testing and risk analysis.

What the labs are proposing

The core idea is straightforward: establish shared protocols for testing frontier AI models before they’re released to the public. That includes independent evaluations, pre-release safety reviews, and standardized risk assessments.

The labs hold different views on how much government involvement is appropriate versus industry self-regulation. OpenAI published a blog post in September 2026 emphasizing its commitment to advancing voluntary AI standards regardless of government intervention, while simultaneously supporting specific legislation in California.

Anthropic has leaned more explicitly toward government partnership. In July 2026, Anthropic’s red team lead publicly advocated for industry-wide safety standards developed in collaboration with government bodies. The company’s Responsible Scaling Policy, or RSP, has been one of the more detailed internal frameworks any lab has published, laying out how Anthropic decides when a model is safe enough to deploy.

OpenAI has its own version called the Preparedness Framework, which similarly attempts to codify how the company evaluates catastrophic risks. Google DeepMind rounds out the trio with its own internal safety research agenda, though it has been somewhat less public about the specifics of its internal protocols compared to its peers.

The Frontier Model Forum and existing infrastructure

This isn’t the first time these companies have tried to coordinate on safety. The Frontier Model Forum, a consortium that includes Anthropic, Google, Microsoft, and OpenAI, was created specifically to address the safety of the most powerful AI models. Its stated goals include developing technical evaluations and funding safety research.

Recent incidents involving AI models, including reported sandbox escapes where models attempted to operate outside their intended constraints, have sharpened the urgency of these discussions.

The report card nobody is bragging about

The Future of Life Institute’s AI Safety Index from summer 2026 is revealing. Anthropic received the highest rating among the three at C+, with a score of 2.66. OpenAI came in at C with a 2.28. Google DeepMind trailed slightly at C with a 2.01.

The ratings suggest that the industry has been retreating from more stringent safety commitments even as the models themselves grow more capable.

What to watch from here

A genuine standards body would need teeth. That could mean mandatory third-party audits before model releases, public disclosure of safety testing results, or agreed-upon capability thresholds that trigger additional scrutiny. Without at least some of these elements, a new standards body risks becoming a more formal version of what already exists: a place where competitors agree in principle and compete in practice.

All three companies are racing to build the most capable models while simultaneously claiming to prioritize safety. Shared standards theoretically solve the competitive tension by ensuring everyone slows down together. Agreeing on who pays the cost of safety, measured in delayed launches and foregone revenue, is where the conversation gets genuinely difficult.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.