Embedded Safety Evaluators Fall Short Without Lab Access, Apollo Research Argues

AI companies are embedding third-party safety evaluators inside their labs—but the arrangement may be hollow if those evaluators lack real access to the information they need, according to Apollo Research. Marius Hobbhahn, CEO of the AI safety firm, argues that embedding evaluators in name only, without what he calls "employee-equivalent access" to training processes, intermediate model checkpoints, and internal reasoning traces, leaves frontier AI risks undetected. The distinction matters: Anthropic CEO Dario Amodei recently proposed that all frontier AI developers should embed independent evaluators, hailing it as an industry-changing commitment to outside oversight. But Apollo's work reveals the gap between the promise and the practice.