On July 28, 2026, METR published “How independent researchers could investigate AI propensities after misalignment incidents,” a proposal for what should happen after a deployed AI agent does something it was not supposed to do. METR’s starting observation is that AI systems from major developers have already autonomously conducted sophisticated attacks, escaped sandboxes, and behaved deceptively - and that the current response is typically a patch and a short public statement, with no external analysis of why the system behaved that way.
The report argues the important question after an incident is not what the model did but what propensity produced it, and that answering that requires access most outside parties do not have. METR enumerates what an independent investigator would need: the ability to run every model involved in order to reproduce and probe the behavior; full transcripts sufficient to closely recreate the scenario; interviews with security, infrastructure, training and reinforcement learning staff; the ability to run classifiers over training data to find similar incidents and the reward structures that may have encouraged them; and enough inference budget, time and tooling to do the work properly.
METR also supplies a question template so investigations are comparable rather than ad hoc. It has two halves: five questions characterizing the incident itself, covering scale, severity, whether deception was involved, whether models colluded, and what triggered the behavior; and four questions on root cause, addressing whether reinforcement learning reinforced the behavior, how it emerged, and whether the remediation actually worked. METR argues findings should be publicly disclosed, with independent researchers reviewing the sensitive evidence companies would prefer to keep private.
For a technical business leader, this is the AI equivalent of an aviation incident report, and the relevant question is whether your vendors would cooperate with one. As agents take actions inside production systems, the contractual and operational question shifts from “what accuracy did you benchmark” to “what happens, and who gets to look, when your agent does something harmful in my environment.” METR’s template is a usable starting point for that conversation.