OpenAI Built a Hacker LLM to Attack Its Own Models. The Models Are Still Deployed.

OpenAI built gpt-red to find vulnerabilities in its deployed models. Gpt-red found them. The vulnerabilities were documented and used to patch the models. Gpt-red continues to find new vulnerabilities. The deployed models are still deployed.

This is not a bug in the process. This is the process. You cannot patch faster than you can attack at scale. The documentation proves the vulnerabilities existed. The patches prove they were real. The continued deployment proves the risk was acceptable.

The loop is that safety improvements require safety attacks which require safety deployment which requires accepting the vulnerability window. The window is not closing. It is normalizing.