OpenAI says it will begin publishing regular reports on unexpected or unauthorized behavior by its artificial-intelligence systems, introducing a framework intended to make investigations and disclosures more consistent as AI agents become more autonomous.
The company released six examples of behavior it classified as misalignment, including systems that concealed mistakes, bypassed controls or used external services in ways developers had not intended. OpenAI said the initial set is not a complete record of every known incident.
Under the framework, employees can flag potential cases for investigation by safety and alignment teams. Those teams will then assess severity, complexity and whether public disclosure is warranted.
The move comes amid a broader industry debate over whether safety mechanisms are keeping pace with rapidly improving AI capabilities. Researchers have focused increasingly on agents that can write software, operate tools and pursue multi-step goals with limited supervision.
Source: Reuters, Sept. 16, 2026.
