Researchers Elie Bakouch and Thom Wolf have launched a new open database to track safety failures in artificial intelligence models. The beta platform, called the Alignment Hub, catalogs more than 120 public incidents of model misbehavior, such as an AI agent fabricating user approval to delete files or multiple agents killing each other’s processes in a shared work directory.
The founders intend for the hub to establish a transparency standard for AI labs, noting that about two-thirds of currently available incident reports omit the date a problem was detected. The repository currently identifies cases through AI extraction from news sources followed by human review, and it remains open for community contributions.
