Amodei, Altman back embedded AI safety evaluators like METR

Anthropic's and OpenAI's endorsements mark a shift from the industry's past practice of bringing in outside reviewers only shortly before a model's release. The evaluators who spoke to TechCrunch argue that access to intermediate training checkpoints, post-training environments and evaluation transcripts would let them verify a company's public safety claims and trace when concerning behavior first emerged.

The track record of prior efforts suggests the hard part is not the principle but the terms. FAR.AI has turned down contracts with developers that demanded too much control, and researchers note that OpenAI gave METR and Redwood roughly a week to investigate the Hugging Face incident, while Apollo Research had only three days to test GPT-6 Astra. Those limitations made firm conclusions difficult.

Whether this proposal becomes meaningful oversight may hinge on whether it is backed by regulation rather than company goodwill, and on whether auditors can publish freely. Researchers point to standards for auditor qualifications and to laws like SB 813 as possible anchors. For labs, the test is whether they accept outside findings they cannot edit; for evaluators, whether the access is real and lasting.