Slowing frontier AI: which governance mechanisms you must now demonstrate per workflow

The call to slow frontier AI means governance must shift from recording to enforcing: independent model evaluations with authority, recorded capacity thresholds that trigger delay and coordinated incident reporting. For your organisation it means you must be able to demonstrate, per sensitive workflow, which frontier models you use and which agreements apply to them.

On 28 July 2026, according to Fortune, more than 1,200 employees of leading AI labs signed the "Pacing the Frontier" statement. They are asking the US government to support an international effort to develop the technical and governance instruments to deliberately slow the frontier of automated AI development. Among the signatories are senior figures from Anthropic, OpenAI, Google DeepMind and Meta; OpenAI and Anthropic endorsed the letter at company level. Why this matters: the question is no longer abstract safety, but concrete governance tools that can moderate the speed of frontier AI.

What exactly did AI workers and leaders ask for in the Pacing the Frontier statement?

According to The Verge, the letter states explicitly that the world currently lacks the technical and governance instruments to deliberately slow frontier-wide progress. DigitalApplied's analysis stresses that the core phrasing does not ask for an immediate pause, but for the building of an option for coordinated slowdown, tested in advance and ready to use if evidence calls for it.

In our assessment this is the pivot point: the signatories diagnose a tooling gap, not an ideological pause. They are asking for infrastructure that makes slowdown possible when evidence justifies it, not for a political vote.

Which concrete governance mechanisms does Anthropic CEO Amodei propose?

On 12 September 2026, Anthropic CEO Dario Amodei, in his essay "We Must Pace the Frontier", as reported by TechCrunch, outlined three strategies that translate the letter into concrete instruments:

• Embedded independent evaluators with employee-level access to frontier systems, to verify companies' pacing and safety commitments.
• Coordinated common safety standards and limits on the pace of uncontrolled AI progress among leading companies in democratic countries.
• Global coordination, including alignment with authoritarian states, on banning limited but clearly dangerous AI applications.

These are governance instruments with teeth: evaluators with access, pace limits tied to benchmarks and international norms. The Cloud Security Alliance places the same campaign in a security governance context and explores which evaluators, thresholds and cross-lab mechanisms are needed to make pacing enforceable.

Which shortcomings in existing AI governance does the pacing call expose?

The diagnosis from the letter - the lack of instruments to deliberately slow frontier progress - points to where current frameworks fall short. General principles such as risk management, transparency and incident reporting already exist, for example in the EU AI Act (Regulation (EU) 2024/1689), but they mainly record; they do not slow anything. In our assessment, these principles must become operational on three points:

• Independent evaluators with access and authority: AI regulators or accredited bodies that can inspect frontier models, run safety tests and, at certain thresholds, recommend or enforce slowdown.
• Recorded capacity and training thresholds: rules that link model size, training method or capabilities (such as autonomous tool use or cyber and bio risk) to "slowdown modes" such as extended evaluation or phased rollout.
• Coordinated monitoring and incident feeds: cross-lab and cross-border mechanisms in which serious incidents automatically feed into pacing decisions, so that slowdown rests on evidence.

This shift ties in with broader expectations around AI oversight under the EU AI Act; we discuss it in our topic hub on AI governance and accountability and in the organisational controls the EU Commission expects.

What must you as an organisation with sensitive workflows now record?

Pacing is not only a matter for labs and regulators. Once frontier models - whether from OpenAI, Anthropic or offered via a vendor such as Microsoft - sit in legal, financial, healthcare or government workflows, you must be able to demonstrate those agreements at workflow level. In our assessment this is the practical translation for professionals who work with confidential information. Record, per sensitive workflow:

1. Which frontier models and agents you use where, and with which version.
2. Which pacing and safety commitments from the vendor apply to those models.
3. Under which conditions you would slow, replace or stop a model.
4. Which logs, evaluations and incident records demonstrate that those controls are being complied with.

On the operational side, it helps to record verification per workflow and to set up red teaming for AI agents in three layers, so that capacity and incident findings actually feed into your decisions.

A verification layer such as IamVera.ai can support this as a visibility layer: it can help show, per workflow, which frontier models and agents are active, which external evaluators or standards they are subject to and which control steps have been recorded. Vera makes control possible by making verification steps, corrections and sources inspectable; it does not assure the correctness of an answer and the final judgement remains with you. The editorial news value lies in the pacing campaign itself, not in any product.

Sources: The article relies on reporting from Fortune, The Verge and TechCrunch and on the Cloud Security Alliance research note and DigitalApplied's analysis of the Pacing the Frontier statement.

← All articles in this topic ← All articles