AEF 1 Evaluator Standard

Anthropic, OpenAI, and xAI have all cosigned the AEF-1 standard, a new baseline for independent third-party AI evaluators. Anthropic is going furthest: permanent, employee-level access for outside evaluators to examine model development, assess alignment, and report incidents. Within hours, Sam Altman and Elon Musk had endorsed it too.

This matters because it changes the shape of the AI safety conversation. For most of this year, the debate has been binary: should frontier labs slow down or not. AEF-1 is the first real answer to a harder question, which is how oversight actually gets built when the companies being overseen are also the ones funding the labs doing the overseeing.

The real test isn't whether evaluators get a badge and a desk. It's whether they can publish uncomfortable findings without the lab that hired them quietly steering the report. Access without independence is theater. Access with it is the beginning of something enterprise buyers, regulators, and employees can actually verify.

If you're evaluating AI vendors for your own stack, this is now a real diligence question, not a hypothetical one.

Do you think third-party access changes how enterprises should evaluate AI vendors, or is this still mostly a PR move until we see an evaluator publish something a lab didn't want out?