OpenAI, Anthropic and Google Discuss Joint AI Safety Standards Amid Antitrust Scrutiny

OpenAI, Anthropic and Google are discussing an industry-led body to develop shared testing and safety standards for advanced AI systems. The talks have continued since July and focus on evaluating increasingly capable models before deployment, according to reporting summarized by multiple outlets.

The discussions bring together companies that compete directly on model performance, cloud distribution and enterprise customers. A shared framework could establish common terminology, evaluation methods and disclosure practices, while leaving each company’s internal safeguards and product decisions under separate control. The effort follows a wider push for voluntary testing agreements and public commitments around frontier AI.

The initiative arrives as AI systems gain broader autonomy and companies prepare increasingly powerful releases. Safety testing is moving from a largely internal responsibility toward a process involving outside evaluators, government institutions and common industry benchmarks.

OpenAI and Anthropic already agreed in 2024 to provide new models to the U.S. AI Safety Institute for testing before and after public release. The National Institute of Standards and Technology said those agreements covered safety research, testing and evaluation, creating a government-backed precedent for structured model review.

A joint standards body would address a gap between voluntary corporate policies and formal regulation. Shared tests might examine dangerous capabilities, cybersecurity risks, misuse pathways and the reliability of safeguards, although the precise mandate and governance structure remain unsettled.

The companies’ earlier participation in the Frontier Model Forum shows that cooperative safety groups are not new. That forum included Anthropic, Google, Microsoft and OpenAI and was designed to promote secure development of frontier models. The current discussions would place greater emphasis on testing standards and could expand coordination among direct competitors.

Cooperation on safety is drawing scrutiny because the same companies control important parts of the AI supply chain. Senator Elizabeth Warren and Senator Ron Wyden opened an investigation into partnerships linking Google and Microsoft with Anthropic and OpenAI, citing concerns that corporate ties could reduce competition, limit customer choice and raise prices.

That concern extends to any shared safety structure that influences model launches, testing requirements or access to technical infrastructure. A common standard could reduce duplication and give regulators clearer benchmarks, but it could also create barriers for smaller developers if compliance requires expensive evaluations or privileged access to testing facilities.

The proposed industry effort would operate alongside government review rather than replace it. The U.S. AI Safety Institute’s agreements with OpenAI and Anthropic established direct channels for pre-release and post-release evaluation, while later policy debates have focused on disclosure, launch timing and federal oversight of frontier systems.

The political environment is also affecting how companies approach safety. Employees from OpenAI and Google DeepMind joined an amicus brief supporting Anthropic in its legal dispute with the Pentagon, illustrating that disagreements over military use and corporate control can cross company lines. OpenAI has separately strengthened its preparedness organization, hiring former Anthropic employee Dylan Scandinaro to lead work on mitigating AI risks.

The immediate test is whether the companies turn discussions into a documented structure with defined membership, independent governance and public evaluation criteria. Without those details, the effort remains a coordination proposal rather than an operating safety regime.

Regulators and developers will watch whether the standards are voluntary, how results are disclosed and whether smaller labs can participate. The arrangement will also face a credibility test: safety rules need enough technical substance to guide deployment decisions without becoming a vehicle for dominant firms to set market terms.