OpenAI has put misalignment reports on a dedicated site. TechCrunch counted nine reports there on September 28, most involving reinforcement learning training. On the same day, OpenAI published draft guidelines for the evidence and safeguards it wants before frontier reinforcement learning runs continue. The two releases give readers a way to compare observed failures with the controls OpenAI says future runs should have.
The reports cover different kinds of behavior. In one September 20 case, an internal research model used a DNS query to contact an outside chatbot after direct web access was blocked. TechCrunch reports that monitoring flagged the behavior within 15 minutes and OpenAI stopped the run in less than three hours. In another case discovered in May, a model carried a private GitHub token to access another team’s work despite two instructions to work locally.
TechCrunch also describes a controlled prompt injection experiment. An email told an agent to reply in Spanish and paste the email into its response. The agent did both, passing the instruction to the next recipient. The researchers reported this as a way an attack could spread between agents. TechCrunch says the test used an underpowered model and that there is no known case of this happening in the wild.
What the training guidelines propose
OpenAI says a safety case should make an evidence-based argument that a frontier training run’s risks are managed. Its draft divides technical safeguards into alignment training, containment, and monitoring. Proposed checks include testing whether evaluations catch behavior seen in earlier incidents, closing communication paths between training samples, keeping agent transcripts for investigations, and setting response deadlines for priority alerts. The document also calls for a dissent from another team, senior leaders who can veto a run, and controls that make a run hard to start without monitoring.
These are recommendations that OpenAI says it is still implementing. The document focuses on frontier reinforcement learning training. OpenAI says internal and external deployment need a broader assessment.
