OpenAI’s Sam Altman Backs Anthropic Plan for Independent Embedded AI Safety Evaluators

Sam Altman, chief executive of OpenAI, posted on X on 12 September 2026 that his company intends to follow Anthropic in granting independent, third-party evaluators the kind of deep internal access normally reserved for employees. The pledge puts two of the world’s most powerful AI laboratories on record as supporting a significant shift in how frontier AI safety is overseen — though critics are already asking what, exactly, these commitments will look like in practice.

The post came in direct response to an essay by Dario Amodei, chief executive of Anthropic, titled “We Must Pace the Frontier”, in which Amodei called for slowing the development of the most advanced AI systems and for embedding independent evaluators inside AI labs on a permanent basis.

Altman wrote: “I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon.”

What “Embedded Evaluators” Actually Means

Amodei’s essay sets out a three-part plan: embedded third-party evaluators inside AI labs; democratic coordination mechanisms; and a global “pacing ladder” intended to slow or sequence the deployment of progressively more capable models.

The embedded evaluator idea is the first step Anthropic has committed to unilaterally. Under its proposal, independent organisations — Amodei cited METR, the Model Evaluation and Threat Research group, as an example — would receive desks, badges, laptops and system permissions comparable to those of internal risk assessment teams. They’d be present during training, able to verify safety practices, report incidents, and assess how well models are aligning with intended behaviour. According to Anthropic’s own description, these evaluators’ contracts would allow them to publish findings without editorial control by the company, subject only to narrow redactions for security, legal and third-party confidentiality reasons.

That’s a meaningful departure from traditional external compliance audits, which tend to be point-in-time exercises rather than continuous, day-to-day scrutiny.

A Rare Moment of Cross-Industry Agreement

Altman isn’t alone. According to BBC reporting and technology news coverage, leaders of at least four major frontier AI laboratories — Anthropic, OpenAI, xAI and Google DeepMind — have publicly expressed support for Amodei’s call to pace the frontier. Elon Musk of xAI and Demis Hassabis of Google DeepMind are among those who have signalled backing for the broad direction.

That’s a rare degree of alignment across organisations that are, in most respects, fierce competitors.

The timing also matters. Earlier in 2026, AVERI — an independent AI safety audit non-profit founded by Miles Brundage, a former OpenAI policy lead — launched with the explicit argument that frontier AI models need independent safety audits rather than industry self-assessment. That initiative added momentum to calls for external oversight that Amodei’s essay has now amplified.

The Gap Between Intent and Implementation

But the pledges, as they stand, are statements of intent.

As of mid-September 2026, neither OpenAI nor Anthropic has published detailed implementation plans, named specific evaluator organisations beyond METR as a cited example, or set start dates. The number of embedded evaluators, the precise scope of their access, and the budgets involved remain unverified, because detailed documentation has not been made public.

Some AI safety researchers and civil society groups welcome the concept while questioning whether evaluators selected and contracted by the companies themselves will have sufficient independence, above all in the absence of any legal obligation or formal public governance structure. Others warn that “pacing the frontier” should not be defined solely by the industry, calling for democratic oversight and statutory regulation rather than voluntary commitments.

Amodei’s essay does include a risk scenario that gives a sense of the urgency he feels: without pacing, he warns, highly capable “agent swarms” of AI systems could potentially scale to take over large parts of the internet within an estimated six to twelve months. That figure is presented as a risk scenario rather than an empirically measured prediction, and should be read accordingly.

Where UK Policy Sits

The UK Government established the AI Safety Institute specifically to independently test and evaluate advanced frontier models, and officials have generally framed independent evaluation as complementary to — rather than a replacement for — statutory oversight. UK regulators are likely to view commitments by Anthropic and OpenAI as a positive signal towards industry self-regulation, while also considering whether formal requirements are eventually needed.

On top of that, the AI Safety Institute’s existence means the UK already has national infrastructure oriented around exactly the kind of independent evaluation that Amodei and Altman are now discussing at the industry level. Whether voluntary lab commitments and government evaluation bodies will be designed to work together — or will develop in parallel — is one of the unanswered questions this debate is raising.

What This Means for Kent Residents

For people in Kent using AI-powered tools through work, the NHS, or local public services, the practical effect of these pledges is not yet clear — OpenAI has said it will adopt embedded evaluators but has not provided timelines or detail, so no immediate changes to products or services are expected. NHS Kent and Medway Integrated Care Board and Kent County Council are both exploring AI-supported tools; if independent evaluation becomes a genuine industry norm rather than a stated aspiration, it could eventually inform the due diligence and procurement standards applied when those organisations assess AI technologies. For now, Altman’s post is a policy signal, not an operational announcement.