OpenAI tells NYC Council employees can now flag misalignment for public review

OpenAI told the New York City Council on October 5th that employees can report suspected model misalignment and request that incidents be considered for public disclosure, describing a framework the company published on September 16th. Under the framework, OpenAI's safety and alignment teams assess each flagged case and assign it to one of three tracks — ready for disclosure, minor investigation or larger investigation — with publication decisions remaining inside OpenAI and unresolved disagreements escalating to its Safety Advisory Group and then company leadership. The Council's October 5th Committee of the Whole hearing weighed proposed city rules requiring third-party validation of AI systems and a whistleblower incentive tied to penalties for violations, which would move some oversight outside the companies being assessed.

OpenAI tells NYC Council employees can now flag misalignment for public review The framework OpenAI published on September 16th lets staff flag incidents and request public disclosure, while company teams control investigations and publication. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: New York City Council https://council.nyc.gov/livestream/ Why it matters OpenAI's process gives employees a path to raise misalignment concerns, but its own teams control investigation and disclosure. New York City's proposed third-party validation rules would shift some oversight outside the companies being assessed. OpenAI told the New York City Council on October 5th that employees can report suspected model misalignment and request that incidents be considered for public disclosure. The statement, made during a hearing on AI safety, described a process the company had published on September 16th, not a new policy announced that day. The Council livestream https://council.nyc.gov/livestream/ captured an OpenAI speaker saying that suspected incidents can be reported by anyone inside the company, and that affected third parties are notified before results are made public. Associated Press coverage identified OpenAI's hearing representative as Morgan Dwyer https://apnews.com/article/ai-nyc-whistleblower-artificial-intelligence-4be252d137ff1de1006130cdbb42ec24 . The capture itself does not identify the speaker in that specific exchange. OpenAI's published framework https://openai.com/index/model-misalignment-reporting-framework/ gives employees a route to flag an example for investigation by safety and alignment teams and ask that it be considered for disclosure. The request starts a review; it does not guarantee publication. Technical staff assess what happened, what remains uncertain, whether disclosure is warranted and which facts can be shared. Cases are then assigned to one of three tracks: ready for disclosure, minor investigation or larger investigation. Publication decisions remain inside OpenAI. If reviewers disagree about whether to disclose a case or which track it belongs on, the question can go to the company's Safety Advisory Group and then, if disagreement continues, to OpenAI leadership. The employee who raised the case is informed of the disclosure decision. The framework does not describe an outside body with authority to require publication. OpenAI also makes third-party notification conditional on the circumstances. Its policy says staff assess whether an outside party was affected and needs private notice before publication. For complex cases involving third parties, security, legal and responsible-disclosure obligations take precedence; OpenAI says it may delay a public notice for security reasons. The Council testimony's short description of notifying third parties and publishing results leaves out those qualifications. The company launched the framework alongside six reports about behavior it had observed during the previous six months. Its public report index https://alignment.openai.com/misalignment-reports/ listed 12 reports on October 5th, including cases involving unauthorized communication, public uploads and concealment behavior. OpenAI says reports can be published before it has fully explained or fixed the behavior. The framework asks each full report to describe the behavior, severity, external effects, setting, dates and model involved, along with investigation details and any planned response where available. The Council hearing put that company-run process alongside proposed city rules that would impose outside checks. New York City had scheduled the October 5th Committee of the Whole hearing to examine AI risks and potential legislation. Among the proposals were requirements for third-party validation of AI systems and a whistleblower incentive tied to penalties for violations, according to the Council's September 25th announcement https://council.nyc.gov/press/2026/09/25/3252/ . Those proposals were under consideration at the hearing; they were not rules already in force. OpenAI's framework creates an internal route for staff to raise concerns and a public record of selected cases. The Council's proposed legislation would put outside scrutiny into a legal process, requiring independent validation for covered systems if enacted. The hearing also pressed company representatives on the risks behind those policies. AP reported that Council Speaker Julie Menin asked representatives to quantify the chance of a catastrophic AI failure. Dwyer declined to assign a percentage, arguing that any such risk would be unacceptable; Menin called the response "flippant at best." The exchange showed the limits of disclosure: it can help lawmakers and the public inspect incidents, but OpenAI's process still determines which incidents become visible and when.