Google, OpenAI and Anthropic plan a frontier AI standards body that would develop common approaches to risk assessment, testing and pre-release review, according to reporting published this week. The proposed organization would operate outside direct government control and focus on the most capable models.

 

The reported plan centers on three functions:

  • Set shared frontier-model safety guidelines
  • Run or coordinate benchmark testing
  • Standardize pre-release risk reviews

 

Google, OpenAI and Anthropic’s Frontier AI Standards Plan

The Information reported that the companies are preparing an organization tentatively called the Standards Authority for Frontier AI, or SAFA. Follow-up coverage by CIO and GovInfoSecurity described it as an independent industry body rather than a new government regulator.

 

The timing is not final. Reports place a possible launch near the end of 2026 or in early 2027, which means the group’s structure, funding and authority could still change before it begins operating.

 

The companies have reportedly discussed leadership roles with figures from technology policy, government and business. Names cited in coverage include Sriram Krishnan, Arati Prabhakar, Condoleezza Rice and David Friedberg, while researchers Beth Barnes and Paul Christiano were mentioned as possible scientific advisers.

 

Those discussions suggest the founders want SAFA to look broader than a three-company committee. Whether the organization can act independently will depend on its charter, voting rules, disclosure practices and ability to publish conclusions that may be uncomfortable for its funders.

 

How SAFA Could Test Frontier AI Models

The reported design traces back to Google DeepMind chief Demis Hassabis’s proposal for an industry-backed organization modeled loosely on a self-regulatory body. The core idea is to create common technical baselines that make results comparable across laboratories.

 

Shared testing could reduce a persistent problem in frontier AI: each developer currently chooses much of its own evaluation framework, release threshold and reporting language. Even when labs publish safety documents, differences in definitions can make direct comparison difficult for regulators, customers and researchers.

 

A standards body could establish repeatable tests for dangerous capabilities, loss of control, cybersecurity misuse and model autonomy. It could also define what evidence a company should produce before releasing a system and how serious incidents should be reported after deployment.

 

The proposal remains distinct from a regulator. Without statutory authority, SAFA could not impose fines or compel participation in the way a government agency can. Its influence would come from industry adoption, public credibility and the willingness of major buyers or governments to treat its findings as meaningful.

 

Funding and Independence Will Define SAFA’s Credibility

Advanced model evaluations require scarce technical talent and expensive computing resources. Hassabis has argued that a capable institute would need substantial financing, likely supplied largely by the industry whose systems it evaluates.

 

That arrangement could make serious testing possible, but it creates a governance challenge. A body funded by frontier labs must show that financial dependence does not allow those companies to soften standards, delay publication or control which models receive scrutiny.

 

Clear conflict-of-interest rules, independent directors and transparent methodologies would therefore be central to credibility. External researchers will also watch whether SAFA receives meaningful access to nonpublic models and internal safety evidence rather than testing only public interfaces.

 

The organization’s membership will matter as well. A body built around three US-based labs could produce useful benchmarks, but wider participation from other developers, independent evaluators and international institutions would make its standards more representative and harder to dismiss as private rulemaking.

 

SAFA Would Join a Crowded AI Oversight System

The proposed body arrives as governments and laboratories debate how to manage increasingly autonomous systems. OpenAI this week called for common measurements, incident-reporting protocols and coordination between national and international standards efforts, especially for systems capable of improving their own capabilities.

 

Existing public institutions would not disappear. In the United States, the Center for AI Standards and Innovation and other federal mechanisms already provide a route for voluntary frontier-model testing, while states continue to develop their own rules. The European Union operates a separate legal framework with binding obligations.

 

SAFA could fill gaps between those systems by offering technical work that multiple governments can reference. It could also move faster than legislation when a new capability or failure mode appears, provided its testing is rigorous and its reporting is sufficiently open.

 

The near-term test is whether the reported coalition becomes a functioning institution with published standards, independent governance and access to leading models. Until those details are released, SAFA should be understood as a significant plan under development, not an established regulator or proof that industry self-governance will work.

 

Further Reading