Governments and OpenAI push independent frontier-AI assessment while common rules remain unsettled

A group of leaders and senior officials from 20 countries, together with European Commission President Ursula von der Leyen, published a joint call on 21 September for tighter oversight of frontier AI. The statement asks companies to use transparent safety protocols with mandatory pre-deployment testing and independent evaluation, asks governments and regional organisations to coordinate common standards and shared reporting of serious safety incidents, and asks UN member states to explore an international institution able to set standards, enable verification and convene states when capability thresholds are crossed.

The signatories include leaders or senior officials from Norway, Finland, Australia, Bahrain, Canada, Denmark, Estonia, Germany, Iceland, Ireland, Kazakhstan, Kenya, Latvia, Moldova, the Netherlands, Singapore, Spain, South Africa, Türkiye and the United Arab Emirates. The text remains open for further endorsements. Its material effect today is political rather than legal: the declaration creates no binding testing requirement, regulator or approval process by itself.

OpenAI’s 21 September standards paper proposes shared measurements of frontier capability progress and research autonomy, human-review triggers and common incident classification and reporting. Reuters independently reported the proposal. OpenAI presents the United States as a preferred leader of that standards effort, but its paper does not establish adoption by governments, standards bodies or rival labs.

Anthropic’s 17 September R&D Automation Index measures a different part of the same governance problem. Anthropic says Claude led 26% of the AI research-and-development work it measured in August and collaborated on more than 90%, with no measured subset fully autonomous. Those are confirmed internal measurements rather than a cross-lab benchmark: Anthropic used its own Claude-based agents and judges, notes there is no common industry methodology, and says independent third-party evaluation is still being set up.

On 22 September, OpenAI said it is committed to supporting independent assessments with deep access across training, evaluation and deployment. It proposes four priority areas: safety cases, critical safeguards, capability and alignment evaluations, and independent investigation of critical misalignment incidents. The company says assessors should be able to challenge assumptions, identify missed risks and reach their own conclusions about safeguards. This is a first-party commitment and proposed assessment model, rather than evidence that every OpenAI system has already undergone such scrutiny.

OpenAI’s proposed principles call for clearly scoped and pre-registered claims, proportionate access within legal, security and intellectual-property constraints, transparent methods and uncertainty, assessor expertise and conflict disclosure, security and confidentiality protections, actionable findings and evidence-based publication. That direction overlaps with the governments’ demand for qualified independent evaluators with sufficient access. Anthropic and Accenture’s separate embedded-evaluator programme remains a company-funded arrangement, so the three efforts should not be treated as one verification system.

On 23 September, OpenAI CEO Sam Altman told the UN Security Council that AI labs should avoid training systems unless they can make a very strong case that the systems can remain under human control. He said OpenAI has slowed development unilaterally before and would do so again when necessary. That is a stated company policy position, rather than a binding obligation or evidence that a particular training run has been halted.

Altman also called for complementary national and international standards covering capability measurement, risk assessment, safeguard sufficiency, meaningful human oversight, rapid incident classification and reporting, and secure channels for governments, critical-infrastructure operators and technical experts to share emerging vulnerabilities. The UN separately confirms that OpenAI and Anthropic executives briefed the Security Council on 23 September. Neither source shows that these proposed mechanisms have been adopted as international rules.

Reuters and TechCrunch reported on 15 September that OpenAI, Anthropic and Google DeepMind had been discussing AI safety across company lines for several weeks, but those talks have produced no public joint charter, common evaluation protocol, standing institution or binding deployment rule. Europe’s role is becoming more concrete through the new declaration and planned frontier-lab discussions, while French and German officials have rejected broad calls to halt AI progress. Nothing reviewed here shows the three labs jointly endorsing the government statement, OpenAI’s standards structure or its new third-party assessment principles.

The evidence remains mixed across independently announced tracks. The government statement is an official multilateral declaration, and OpenAI’s assessment commitment and Anthropic’s measurements are established as first-party evidence, but there is still no adopted common frontier-AI standard. A formal UN or recognised standards process, binding national or EU rules, a published cross-lab protocol, or independent assessments demonstrating comparable methods across labs would materially strengthen the evidence for a unified governance framework.