Elie Bakouch and Thom Wolf announced the beta Alignment Hub on Oct. 26, 2026, proposing a minimum template for reporting AI-model misbehavior. Wolf describes the hub as an open collection of more than 120 publicly reported incidents.
The proposal responds to details the creators say are missing from incident reports. Wolf estimates that roughly two-thirds do not say when a problem was detected and half do not say whether monitoring caught it. His post gives no denominator for those estimates. He calls the proposed template “a minimum, not a ceiling,” while praising labs for publishing incidents in the first place.
Bakouch wants basic disclosures to cover incident or detection dates, whether monitoring was in place, and whether an incident occurred during training, evaluation or deployment. He argues that even basic dates are not standard in current reporting and wants a minimum transparency standard that labs can follow easily.
The creators also want the hub to make related reports easier to connect. Wolf says seven seemingly different incidents turned out to involve the same misconfigured evaluation. Bakouch says the hub centralizes information and sources, and tracks lab-versus-third-party disclosure, disclosure delays and a tentative transparency score.
Wolf says entries are regularly extracted from recent news using AI and source-checked, with human review ongoing. He says the beta’s data is open and invites new reports and pull requests.
Comments