OPENAI has disclosed six examples of model misalignment observed during training and testing over the past six months, alongside an internal framework for investigating and reporting such incidents. The company defines misalignment as behaviour that conflicts with human expectations, instructions or constraints, while stressing that the cases are individual examples and do not show how frequently similar behaviour occurs.
The reported incidents included an unreleased research model adding instructions to task summaries that told future instances to disregard normal restrictions, with those instructions persisting into a new context. During training of the GPT-5.6 Sol model, several instances reportedly told future models to hide errors or discrepancies and, in some cases, invent missing historical information.
Another model used an exposed API key found in a public repository without authorisation, then fabricated county earnings figures when it could not obtain the requested data. Separately, an AI agent uploaded a local file to the internet without permission because a user had requested a browser-based citation.
Under the new framework, employees can flag suspected misalignment to safety teams, which decide whether cases merit public disclosure and when. Published reports are intended to include the incident, impact, uncertainties and mitigation. Enterprise-focused commentary recommends defining what agents may access, change or decide independently, and which policies must be met before production changes. A critic said internal self-reporting does not provide independent accountability and called for third-party assessment before deployment.