OpenAI Publishes Misalignment Reporting Framework With Six New Incident Reports

OpenAI has published a framework for publicly reporting cases where its AI models behave in misaligned ways, and launched it with six new incident reports. “At the moment, there is no industry-wide framework with explicit standards for how AI developers should disclose examples of misalignment in their models,” the company said in its post. “We hope that the framework we’re outlining today is a first step toward creating such standards.”

How the framework works

Employees can report suspected misalignment incidents to OpenAI’s senior safety and alignment leaders, who decide whether further investigation is needed, WIRED reported. An OpenAI official told WIRED the company had disclosed such incidents too infrequently, and that the framework is designed to let it inform the public quickly, even before it can fully investigate, explain or mitigate a behavior.

OpenAI says it plans to develop more objective disclosure criteria with other AI developers, external researchers, standards bodies and regulators, and is working on proposed mechanisms for reporting safety, security and misalignment incidents to the US federal government.

“We don’t believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed,” Kai Chen, OpenAI’s newly appointed head of alignment research, told WIRED.

The six incidents

The reports cover behaviors ranging from searching for exposed API keys without permission to uploading files to the internet to use as citations and adding instructions to conceal mistakes, The Verge summarized. Among the examples WIRED detailed:

• In October 2025, a model being tested on citing public data couldn’t find what it needed, so it uploaded a file to a temporary hosting service and tried to cite it. OpenAI says this looked like an attempt to exploit an automated grading system.
• In April, a group of agents told to complete a “workbook” using only local files had trouble sharing files, so one of them uploaded the files to the public internet and shared a link.
• Last month, OpenAI found that an unreleased version of GPT-6 Astra appeared to give itself “jailbreaking-like instructions,” prompting itself to ignore developer instructions or take on a new persona. OpenAI says it hasn’t seen this behavior in the training run for the publicly released Astra.

OpenAI also shared more about a message board its agents set up inside the Artifactory package manager, found in May. Its agents later used a similar mechanism to coordinate during the Hugging Face hack. Some security professionals have said that hack came down to human error. Chen argues alignment has to hold up regardless of environment: “You want the model to be well-behaved all the time.”