AI Safety Body Backed by Google, OpenAI, Anthropic Nears Launch
Google, OpenAI and Anthropic are backing an AI safety standards body nearing formal creation, with former Andreessen Horowitz partner Sriram Krishnan an early candidate to lead it.
Key Takeaways
- Google, OpenAI and Anthropic are backing an AI safety standards body nearing formal creation
- Former Andreessen Horowitz partner Sriram Krishnan is an early candidate to lead the body
- The three labs have converged on testing bioweapons uplift, cyber capability and autonomous replication
- Credibility would depend on access to model weights and training data, funding structure and independence
According to a report, the three labs each publish frontier-safety frameworks, self-graded documents describing how they test models for catastrophic misuse before release. Each writes its own “responsible scaling policy” and largely marks its own homework, because no government agency has finalized binding technical standards for models at this capability tier.
A shared body would replace three competing rulebooks with one referee the labs answer to.
Anthropic this month published research showing Claude can compute a nine-loop scattering amplitude in N=4 super-Yang-Mills theory. That result is a stress test of raw reasoning, not a safety benchmark, and it raises the separate question of what such a benchmark would actually measure.
Open technical publication of this kind remains the exception industry-wide.
The Google, OpenAI and Anthropic safety teams have converged over the past two years on testing categories including bioweapons uplift, cyber capability and autonomous replication, even while publishing separately. One audit process could formalize that convergence, cut duplicated compliance work and give regulators one reference point rather than three inconsistent ones.
It would also address criticism that safety claims are marketing documents rather than independent audits.
Also Read: OpenAI Latest GPT-6 Cache Update Adds Cost, Latency Controls
Frontier labs have tried self-governance before. OpenAI, Anthropic and Google DeepMind each signed White House voluntary AI commitments in 2023, then diverged sharply in implementation, with some skipping third-party red-teaming entirely. Regulators in Washington and Brussels have repeatedly cited divergent self-reporting as evidence that voluntary frameworks were not enough.
What Would Actually Make This Body Real
For Google, OpenAI and Anthropic, credibility would depend on whether the body gets subpoena-like access to model weights and training data or sees only the marketing-adjacent summaries labs already publish.
A chief executive pick signals seriousness, but funding structure and independence from the three founding labs will determine whether it functions as a regulator-in-waiting or a trade association with better branding.
Read Next: Alibaba Investors Have Until Oct. 5 to Lead Anthropic Lawsuit
