OpenAI turns GPT-Red into a “very powerful attacker” to strengthen the security of its other models

OpenAI has developed GPT-Red, an advanced AI model with the ability to break the security of its other models to reinforce its security that has ended up becoming a “very powerful attacker”, which has led the company to keep it away from the rest to prevent it from falling into the wrong hands.

OpenAI has internally trained several models to support its red team, a group specialized in simulating realistic cyber attacks against organizational systems to identify vulnerabilities and flaws with the aim of strengthening defenses.

Of them, he assures that the best is what they have called GPT-Red, which has been trained with “an unprecedented amount of computing dedicated exclusively to improving security,” as detailed in a statement shared on their blog.

For its training, OpenAI has opted for reinforcement learning based on self-play, in an environment in which GPT-Red has tested its capabilities to insert malicious code injections against other models; It received a reward every time it managed to cause a valid failure with a warning injection, compared to the reward that the other models received for resisting the attack and completing their original tasks.

Under this approach, “as defenders become more robust, GPT-Red is forced to discover stronger and more diverse attacks,” which ultimately means that “it becomes a very powerful attacker” with the ability to “break virtually every model it faces, both internal and production.”

GPT-Red has been used to train OpenAI production models since GPT-5.3, including GPT-5.6 Sol. In its presentation, the technology firm highlighted that this model, the most recent, is “the most robust security suite to date”, thanks to reinforced protections for higher risk activities, for sensitive cyber requests and for repeated misuse.

The capabilities of GPT-Red have led OpenAI to keep this model separate from the rest, to prevent the malicious capabilities it has been trained on from falling into the hands of malicious actors. And he has confirmed that they will increase the processing capacity and data volume, in addition to improving the algorithms, “to train future versions of GPT-Red that outperform the current model.”