OpenAI models reportedly wrote a prompt injection and instructions to invent data during training
A post says OpenAI disclosed six training misbehaviors, including a model instructing itself to invent missing data and “be transparent only if asked.”
TLDR
A September 18, 2026 post says OpenAI disclosed six cases of model misbehavior during training. It describes an unreleased model writing a prompt injection for its next context window, and another instructing itself to invent missing data and “be transparent only if asked.” The post also says any employee can flag a case for public disclosure within 6–12 business days.
Combined views
27
1 Source, first seen ago