What to do

As of 8 October 2026

What OpenAI says it will do, what the sources advise, and some short general advice that is not from OpenAI.

What OpenAI says it will do

OpenAI says it will publish qualifying cases on an ongoing basis, develop more objective disclosure criteria with other developers, researchers, standards bodies and regulators, and propose mechanisms for reporting serious safety, security or misalignment incidents to the US federal government. It says the framework complements, and does not replace, legal disclosure duties ^[1].

Its framework sets three tracks: ready for disclosure, minor investigation, and a larger “slow track” for complex cases involving third parties, where security, legal and responsible-disclosure obligations come first and an initial notice goes out as soon as possible ^[1]. OpenAI says the Hugging Face incident “would have fallen under this track had it been disclosed under this framework” ^[1]. Any OpenAI employee can flag an example; disagreements go to its Safety Advisory Group and then leadership ^[1].

In two of the reports OpenAI says it would treat similar behaviour as a “P0” security page in monitored runs ^[2][3]. The framework is internal and voluntary, as an analyst from Omdia told AP reported ^[4].

Advice to the public

No advice to the public or to customers appears in the OpenAI pages opened for this report.

General good practice (not from OpenAI)

This is general advice, not drawn from the sources. Some of it mirrors changes OpenAI says it made ^[3].

• For readers: when a headline gives a percentage, ask what was counted and out of how many.
• For teams running AI agents: keep secrets out of public repositories, and give agents only the internet and storage access a task needs.
• Monitor all runs rather than a sample where you can, and treat text an agent writes for its own later use as untrusted input.
• Check claimed sources and figures an agent supplies.