OpenAI, Google, and Anthropic have reportedly been holding quiet talks about building a shared industry body to measure the safety of advanced AI models. The idea draws on FINRA, the self-regulatory organization that oversees US securities firms, and aims to give the industry a common yardstick for testing increasingly powerful AI systems. Fierce rivals though they are, the three companies appear to agree that rapid gains in AI capability call for some coordinated ground rules.

Hassabis's FINRA-Style Proposal

Google DeepMind CEO Demis Hassabis publicly called for a Frontier AI Standards Body in July. The model he has in mind is FINRA, the self-regulatory body that oversees US brokerages. Under his proposal, companies with frontier-level models would voluntarily submit them for roughly 30 days of pre-release review by independent experts. That review would cover dangerous capabilities such as cyberattack potential, misuse for bioweapons, and deceptive behavior. Hassabis has said he would like the organization up and running before the end of 2026, with a possible role for government involvement further down the road.

Working-Group Talks Have Continued Since July

Following that proposal, representatives from OpenAI, Google, and Anthropic have reportedly held working-group discussions since July, with talks continuing as recently as September. No formal organization has launched yet, and the rules and governance structure remain unsettled. Still, the fact that three fierce competitors keep coming back to the same table is starting to shift how the industry reads the moment. Anthropic CEO Dario Amodei also published an essay in September arguing that the pace of AI capability growth should be deliberately slowed, calling for ongoing scrutiny by independent evaluators and closer cooperation among developers. OpenAI CEO Sam Altman, speaking at an internal town hall, voiced support for an industry-wide testing and audit framework as well.

Three Companies, Three Different Yardsticks

Each company already runs its own safety framework, but the details diverge. OpenAI uses its Preparedness Framework, which sorts risk by capability thresholds. Google DeepMind relies on Critical Capability Levels with built-in early-warning triggers. Anthropic's Responsible Scaling Policy ties capability and usage thresholds to escalating mitigations. The goals overlap, but the terminology and measurement methods don't line up cleanly, which makes it hard for outsiders to compare claims when one company says a model has "crossed a dangerous capability threshold." A shared body could potentially standardize terminology and set a minimum bar for testing across companies.

Independence and Cost Remain Sticking Points

Not everyone is convinced. The biggest concern is independence: if the companies building AI also run the body that grades them, the result could end up as little more than an industry club handing out gold stars. Hassabis's original proposal called for a board with substantial independent representation, including technical experts, government voices, and open-source advocates, but questions about who appoints its leadership and who funds its operations remain unanswered. Testing costs are also expected to be significant, raising fears that the requirement could become a barrier for startups and open-source developers with fewer resources. Altman has floated an even bigger ambition, suggesting the US and China could eventually share testing standards, though that kind of government-to-government agreement faces a much higher bar.

Will It Spread, or Stay Limited to Three Companies?

Airlines compete under a shared set of safety rules, and banks compete under common financial regulation; whether AI can pull off the same balance of rivalry and shared standards is now the open question. If the three companies' cooperation works, other developers such as Microsoft, Meta, and xAI could face pressure to join, and regulators around the world could use the resulting evaluation methods as a reference point. If it doesn't work, the effort risks becoming little more than a symbolic gesture.

Summary

OpenAI, Google, and Anthropic are discussing a shared AI safety standards body built on the FINRA-style framework that Google DeepMind's Demis Hassabis proposed in July. No formal launch has happened yet, and questions about independence and who bears the cost of testing remain unresolved. Even so, the fact that three fiercely competing companies are talking safety standards at the same table marks a new phase in AI governance.

Image for illustration purposes only.