Anthropic and OpenAI Want Embedded Safety Evaluators — but Can They Stay Independent?

Anthropic and OpenAI want independent safety evaluators working inside their own labs, according to a TechCrunch report published on September 16, 2026. The arrangement would give outside researchers a level of access to frontier models that external auditors have long asked for and rarely received.

Under the proposals described in the report, evaluators would sit inside the labs rather than reviewing models from the outside, examining behavior and risk while the systems are still in development. Supporters argue that proximity is what makes early detection possible, letting reviewers catch problems before a model reaches millions of users.

Researchers quoted in the coverage welcomed the unprecedented access, but their support came with conditions. They warn that meaningful oversight requires transparency about methods and findings, genuine independence from the labs being assessed, and eventually formal regulation that does not rest on corporate goodwill.

The central question is structural. If an evaluator sits inside a company and depends on that company's cooperation, publishing conclusions the company dislikes becomes difficult. Without guarantees on publication rights, funding and tenure, access can become proximity without accountability.

The report frames the debate as a test of how the industry's safety commitments hold up as competition intensifies. Anthropic and OpenAI have both positioned themselves as leaders on safety, and both now face the harder question of whether they will accept scrutiny they do not control.

For regulators, the model is attractive because it is faster than legislation and cheaper than building a state evaluation apparatus. For critics, that is exactly the problem: oversight designed and hosted by the organizations it oversees rarely survives contact with commercial pressure.

The signal matters beyond the two companies. Frontier evaluation is becoming the main mechanism by which governments, enterprises and the public learn what new models can do before they are deployed, so the credibility of that mechanism determines how much the rest of the system can trust it.

What to watch next is concrete: whether embedded evaluators get unfettered model access, whether their findings are published in full, who controls their budgets and hiring, and whether lawmakers turn the arrangement into a legal requirement rather than a voluntary one.