OpenAI develops GPT-Red, offensive AI for internal security testing
OpenAI developed GPT-Red, an AI model intentionally trained to break the security of its other models through prompt injection attacks, as part of internal red team testing. The model became so effective that OpenAI isolated it to prevent its offensive capabilities from being misused by malicious actors.
What happened
OpenAI developed GPT-Red, an advanced AI model trained with unprecedented computational resources to break the security of its other models.
The model was trained using reinforcement learning based on self-play, in which GPT-Red competed against other models by attempting to insert malicious code injections through prompt-based attacks, receiving rewards for successful exploits while defensive models were rewarded for resisting attacks.
GPT-Red has been used to train OpenAI's production models starting from GPT-5.3 through GPT-5.6 Sol, and according to OpenAI it has become capable of breaking practically all models it faces, both internal and production.
Reported impact
OpenAI has kept GPT-Red isolated from other systems to prevent its offensive capabilities from falling into the hands of malicious actors.
The company confirmed it will increase processing capacity and data volume to train future versions of GPT-Red that surpass the current model.
Governance controls
OpenAI separated GPT-Red from the rest of its infrastructure to contain the risk of misuse of its hacking capabilities.
The company highlighted that its most recent model, GPT-5.6 Sol, includes the most robust security protections to date, with enhanced safeguards for high-risk activities, sensitive cybersecurity requests, and repeated misuse.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.