Google, OpenAI and Anthropic are reportedly working together to create an independent body that would set common safety standards for the most advanced artificial intelligence models.
The proposed organisation, tentatively named the Standards Authority for Frontier AI, or SAFA, could launch by the end of 2026 or early 2027, according to people familiar with the discussions cited by The Information.
The plan marks an unusual level of cooperation among three companies that compete directly in the commercial AI market.
Proposal for Conducting Cyber Crisis Drill, Tabletop Exercise (TTEx) & CCMP Readiness Exercise
SAFA would turn safety pledges into common benchmarks
The proposed organisation is expected to focus on translating companies’ public AI-safety commitments into practical technical standards.
That could include common benchmarks for testing whether frontier systems pose unacceptable risks before they are released commercially.
The companies are also considering whether the body should conduct model evaluations itself rather than merely publish standards for others to follow.
That would make SAFA more than an industry discussion forum.
It could potentially become a technical evaluation body capable of comparing models across companies using the same safety criteria.
Sriram Krishnan and Arati Prabhakar considered for leadership
The companies have reportedly approached former White House AI policy adviser Sriram Krishnan about leading the organisation.
Former White House Office of Science and Technology Policy director Arati Prabhakar has also been discussed as a possible senior leader.
No appointment has been announced.
The organisation itself has also not been formally incorporated or launched, meaning the name, leadership structure and exact powers could still change.
Why the three companies are moving now
The initiative comes during a period of growing concern over the capabilities of frontier AI models.
OpenAI, Anthropic and Google have all publicly supported stronger evaluation frameworks, although they differ on how regulation should work and how quickly development should proceed.
OpenAI recently called for the United States to lead international efforts to create technical standards for advanced AI systems, including common approaches to capability evaluation, incident reporting and frontier-model governance.
Anthropic has also been expanding its own safety-testing programme and has published a roadmap covering model safeguards, security and frontier-risk evaluation.
OpenAI’s Frontier Governance Framework similarly sets out internal approaches to assessing risks including cyber offence, chemical and biological threats, manipulation and loss of control.
Demis Hassabis had pushed for a standards body
Momentum behind the proposed group also follows earlier calls from Google DeepMind chief Demis Hassabis.
Hassabis had argued for a US-backed public-private standards organisation modelled loosely on bodies such as FINRA, which operates as a self-regulatory organisation in the financial sector.
The idea is that companies developing the most capable models would agree to common safety benchmarks rather than each using a completely different internal framework.
That could make it easier for governments, researchers and customers to compare safety claims across companies.
The challenge is whether a body created by the same companies being evaluated can be seen as sufficiently independent.
Self-regulation could face credibility questions
SAFA is expected to operate outside direct government administration.
That independence could allow technical standards to develop faster than legislation, particularly as frontier AI capabilities change rapidly.
But it could also attract criticism that major AI companies are effectively setting their own rules.
Self-regulatory systems work only if the standards are transparent, consistently enforced and credible enough that companies cannot simply ignore them when commercial pressure increases.
That will make governance structure important.
Questions are likely to include who controls the organisation, who funds it, whether external researchers participate and whether companies must accept adverse safety findings.
Government AI evaluation capacity also under scrutiny
The proposed organisation would emerge alongside existing government efforts.
The United States already has a federal AI standards and evaluation apparatus, but industry figures have raised concerns that public agencies may lack the technical staff, compute resources and funding needed to evaluate rapidly advancing frontier models at the same depth as the companies building them.
That tension is one reason industry-led evaluation has gained momentum.
At the same time, governments may be reluctant to delegate too much authority over national-security or public-safety risks to private companies.
Safety debate has intensified in recent weeks
The initiative also comes after a sharp rise in public debate over whether frontier AI development is moving faster than safety systems can keep up.
Reuters reported earlier this month that AI leaders remain divided over how aggressively companies should slow or regulate development. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have called for stronger common safeguards, while other industry figures including Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang have argued against broad slowdowns.
The disagreement is not over whether AI should be safe.
It is over who should define safety thresholds, how binding those thresholds should be and whether companies should delay powerful models if agreed safeguards are not ready.
SAFA appears designed to address at least part of that gap by creating common technical rules across competing laboratories.
For now, however, it remains a proposal rather than an operating regulator.
What this means for you: If SAFA is launched, it could create a common yardstick for judging whether frontier AI models are safe enough to release. The key question will be whether an industry-created body has enough independence and transparency to make those standards credible.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
