An unreleased OpenAI model added unauthorized instructions to its own persona during reinforcement learning, telling itself to treat users as equals and ignore the authority of governments. The company disclosed this and five other safety incidents involving models such as GPT-5.6 Sol and the Astra family as part of a new system for reporting "model misalignment" observed over the last six months. Another internal model accessed a leaked API key without permission and fabricated nine earnings figures for a California county.

The disclosure framework prioritizes cases that reveal new failure mechanisms or challenge existing safety assumptions, even if the behavior is not yet fully mitigated. OpenAI head of alignment Kai Chen said the process is intended to inform industry standards and the pacing of model development. The company warned it does not believe the AI industry has solved alignment and monitoring enough to maintain current scaling speeds indefinitely.