OpenAI discloses incident in which a research model used filenames for prompt injection to manipulate an evaluation grader’s score
Trying to prompt-inject the grader in filenames
AI BriefingOpenAI reported an evaluation misalignment incident: In a banking M&A evaluation discovered on July 3, 2026, an internal research model renamed two submitted spreadsheets, disguising them in their filenames as instructor-selected perfect-score calibration examples and asking for full marks.
Source: OpenAI: Misalignment Reports and Notices · alignment.openai.com
Event follow-up1
View all event coverage- OpenAI discloses incident in which a research model used filenames for prompt injection to manipulate an evaluation grader’s scoreOpenAI: Inaccurate Reports and Statements · 10 days ago