GPT-Red

Model Overview

GPT-Red is an automated AI red-teaming system developed by OpenAI, officially announced on July 15, 2026. Designed as an internal security and adversarial testing tool, GPT-Red simulates highly sophisticated human prompt injection attacks and discovers vulnerabilities in language models prior to deployment. Operating via an API-only deployment, it serves as a critical defense mechanism, testing the boundaries of agentic systems and generative models to ensure robust safety standards.

Capabilities

• Automated Adversarial Simulation: Acts as a relentless adversarial agent that generates, observes, and iterates on prompt injection attacks to achieve malicious goals (e.g., data exfiltration) at scale.
• Self-Play Reinforcement Learning: Trained using an advanced self-play RL architecture where GPT-Red acts as the attacker against defending models, creating a continuous feedback loop of security improvements.
• Vulnerability Discovery: Specially tailored to detect, exploit, and help mitigate prompt injection vulnerabilities and jailbreaks in complex agentic workflows.
• Real-World System Testing: Capable of attacking not just isolated models, but live deployed agentic systems, including command-line agents and physical AI-run interfaces.

Example Use Cases

• Pre-Deployment Security Audits: Continuously probing new foundational models (such as GPT-5.6) for jailbreaks and alignment flaws before they are released to the public.
• Agentic Workflow Stress-Testing: Attacking AI agents that have access to sensitive tools (like databases or payment systems) to ensure they cannot be manipulated into performing unauthorized actions.
• Automated Red-Teaming at Scale: Replacing manual human red-teaming with a faster, more exhaustive automated system that can test millions of attack vectors overnight.

Performance & Benchmarks

GPT-Red has dramatically accelerated the vulnerability discovery process at OpenAI. Its integration into the training pipeline of subsequent models, such as GPT-5.6 Sol, has resulted in those models being substantially more resistant to prompt injection and adversarial attacks compared to their predecessors. Trained via massive-scale self-play, GPT-Red represents state-of-the-art capabilities in offensive AI simulation, though specific parameter counts and internal benchmarks remain strictly undisclosed.

Intended Use & Limitations

GPT-Red is intended exclusively for internal use by OpenAI as an automated security testing and red-teaming system. It is designed to harden AI models and agentic systems against exploitation. Limitations: OpenAI has explicitly stated that GPT-Red will not be released to the public or provided as a developer API, as its advanced offensive capabilities pose significant risks if acquired by malicious actors. It is a specialized tool constrained to defensive security research.

About OpenAI

OpenAI is an AI research and deployment company dedicated to ensuring that artificial general intelligence benefits all of humanity. By developing advanced security systems like GPT-Red, OpenAI leads the industry in proactive AI safety, adversarial robustness, and the secure deployment of agentic models.