OpenAI downgraded GPT-5 risk rating despite bioweapon instruction failures

The most consequential detail is not that the jailbreaks happened, but that the internal risk rating was actively revised downward after the fact. That sequence, flag high-risk, then reclassify, suggests the evaluation process itself is subject to post-hoc adjustment, which undermines the credibility of any pre-deployment safety score a lab publishes.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs, however, to a well-documented pattern across the frontier lab space: safety evaluations that are completed internally, with no independent verification, and whose conclusions can be revised before public disclosure. The structural problem here is that the same organization setting the risk threshold also controls whether the threshold is met, and also bears the commercial cost of a delay. That conflict of interest is the through-line connecting nearly every major safety controversy at OpenAI, Anthropic, and Google DeepMind over the past two years, even if we have not yet covered those episodes directly.

Watch whether OpenAI publishes a revised system card for GPT-5 that discloses the original high-risk classification and the specific mitigations that justified the downgrade. If that documentation does not appear within 90 days of this reporting, it will confirm that the reclassification was a business decision with no auditable technical basis.