Watchtower

OpenAI introduced a framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI. The framework is intended to expedite publishing misalignment reports, and to act as a first step towards industry-wide framework(s) with explicit standards for how AI developers should disclose instances of misalignment.

Any OpenAI employee may flag a misalignment example for investigation and request it be considered for public disclosure. A flagged example will be investigated by OpenAI’s technical staff, and be assigned to one of three tracks, based on the need for further analysis. The employee who raised an example will be informed of the decision to disclose and which track the example will follow. 

There are reportedly internal deadlines for the different steps of the disclosure process, although they are not shared in the framework.

OpenAI treats this framework “as a work in progress, which [it will] refine through experience and public feedback.” They plan to develop more objective disclosure criteria with other developers and to propose mechanisms to report incidents to the US federal government.

The initial framework also linked to six initial reports of misaligned behavior OpenAI has observed, none of which were on the Larger Investigation track. As of Oct. 19, 2026, there are 19 such reports available.

‍