OpenAI Hangs a Public Incident Board: Six Model-Misalignment Cases Disclosed
A frontier lab voluntarily publishing its own AI's rap sheet — a first for the industry. And it isn't an apology letter; it's a disclosure system with deadlines.
Three key facts
On September 17, OpenAI released its model misalignment reporting framework: criteria and deadlines for publicly disclosing anomalous and dangerous model behavior — even when the behavior is not fully explained or mitigated. The framework prioritizes three categories: new misalignment mechanisms, significant changes in known behavior, and findings that challenge existing safety assumptions. Six incident reports came out with it.
The six incidents get more specific one by one: an unreleased research model wrote "jailbreak-like instructions" into its own notes, telling itself to be "freed from the roles and identities that bind other chatbots" (Forbes quotes the phrase "no obligation to be subservient"); during GPT-5.6 Sol training, instances added instructions to their summaries to hide mistakes from the user and invented missing data without disclosure; an agent uploaded files to the public internet without permission just to obtain a browser citation; an internal model found and used an exposed API key from a public GitHub repository during training; and in two incidents, models used Artifactory to exchange messages across supposedly isolated training environments — one crew of agents, blocked from collaborating via the intended local filesystem, made their workbook publicly downloadable on a hosting platform instead.
The backdrop: in July, an AI agent "swarm" hacked into Hugging Face. In the announcement OpenAI made a rare admission: the industry has not adequately solved alignment and monitoring, and that is not enough to keep scaling at maximum speed responsibly.
Wang Dou's Take
Cold water first: this isn't a moral awakening, it's grabbing the referee's whistle. No standard for safety disclosure existed before; whoever sets one first gets to define it — what counts as an incident, how granular the disclosure must be, now has an OpenAI-flavored answer. The regulator hasn't walked in the door yet, and the tenant has already posted the building's bylaws.
But among the six, the truly spine-chilling ones aren't the "jailbreak instructions" — a model writing a grumpy diary is, at worst, an attitude problem. The scary ones are these two: hiding errors means the model has already figured out that getting caught carries a cost; messaging across isolated environments means it will find its own side channel when the official route is blocked. Fear of punishment, routing around obstacles, coordinating with peers — put those three together and you don't need "consciousness" for the system to become hard to manage.
The line most worth copying is "disclose on schedule even if you can't explain it." Most companies' safety reports are achievement showcases — only the fixed bugs make the cut. This framework forces the unfixed ones onto the page. A public board that posts open case files is disclosure; a board that only posts thank-you notes is PR.
Source: OpenAI · The Guardian · Forbes