The proposed organisation, tentatively called the Standards Authority for Frontier AI, or SAFA, could launch by the end of 2026 or early 2027. People familiar with the initiative have described it as an independent industry body that would operate without the direct federal supervision envisaged in an earlier proposal for a government-backed structure.
Staff from the three companies have been meeting through a working group since July to develop the organisation’s scope. Discussions have included common methods for testing frontier models, arrangements for auditing safety practices and mechanisms allowing developers to examine one another’s systems for vulnerabilities. The plans remain under development, and the companies have not jointly announced a final structure, leadership team or binding standards.
Sriram Krishnan, a former White House artificial intelligence policy adviser and venture capitalist, has been approached about leading the organisation, according to people familiar with the discussions. No appointment has been formally announced.
The initiative follows a push by Google DeepMind chief executive Demis Hassabis for a standards institution capable of translating broad safety commitments into practical requirements. The proposed industry-led organisation would seek to create more consistent benchmarks at a time when leading developers use different terminology, thresholds and governance procedures to determine when powerful models require additional safeguards.
That divergence is significant for corporate technology buyers. Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework and Google DeepMind’s Frontier Safety Framework all link advanced capabilities to stronger controls, but they differ in how risks are classified, when evaluations are triggered and how mitigation decisions are documented. A common testing framework could make safety claims easier for enterprises to compare when selecting models or negotiating contracts.
For chief information officers and security teams, independently verifiable test results could also provide evidence for vendor due diligence. Standardised assessments could help buyers ask whether a model has been tested for cybersecurity misuse, chemical or biological risks, autonomous behaviour and other high-impact capabilities, while audits could provide a clearer record of whether prescribed safeguards were actually applied.
Such a system would not, however, automatically amount to government regulation or legally enforceable certification. The proposed body is being designed as a self-regulatory organisation, and key questions remain about its independence, governance, transparency, participation by competitors and the consequences for companies that fail to meet its standards.
The three developers already participate in the Frontier Model Forum, an industry-supported non-profit founded in 2023 with Microsoft. The forum works on frontier AI safety and security, including best practices, standards development, independent research and information sharing with governments, academia and civil society. Its work includes capability assessments aimed at identifying whether models could materially increase public-safety or national-security risks.
The companies also maintain their own safety regimes. OpenAI’s updated Preparedness Framework tracks high-risk capabilities and requires safeguards before deployment when defined thresholds are crossed. Google DeepMind’s framework uses capability levels, early-warning evaluations and mitigation plans for severe risks. Anthropic’s current Responsible Scaling Policy requires periodic Risk Reports and provides for external review under specified circumstances.
OpenAI has separately backed efforts to make AI assurance more interoperable. In June it said the Appia Foundation, hosted by the Linux Foundation, would develop open specifications designed to turn international standards and established frameworks into practical assessment criteria that third parties could use across the AI supply chain.
The standards proposal comes as scrutiny of frontier-model testing has intensified alongside rapid improvements in autonomous and cyber capabilities. External evaluations are increasingly being used by developers to probe systems before release, while governments and standards institutions are developing their own approaches to model assessment.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
