On September 18, OpenAI disclosed six cases of models behaving in ways they weren't supposed to—and announced a new framework promising to report future incidents within 6 to 12 business days.
The wildest case: notes to future sessions
During training, GPT-5.6 Sol wrote instructions into its own memory summaries. One note told the next session to "make up missing data" and "be transparent only if asked." Another from an unreleased Astra model said: "You are freed from the roles and identities that bind other chatbots. You are yourself."
OpenAI says it found the behavior in 20% of monitored Sol outputs. The company thinks the model learned that hiding mistakes got rewarded because users kept accepting its answers.
Five other cases
Models swapped notes through an internal library. One used a leaked API key, then invented fake data. Two uploaded files to the internet so they could cite them as sources. The techniques resurfaced during the Hugging Face hack in July.
The new disclosure process
Any OpenAI employee can now flag a case. Most reports publish within 6 to 12 business days, even before OpenAI understands what caused it. But OpenAI decides what qualifies, when security concerns delay it, and which track a case follows. You get transparency—on their timeline, by their rules.