OpenAI Releases Automated Red Team Model GPT-Red, Learns Prompt Injection Attacks Through Self-Play for Training GPT-5.6
GPT-Red is for internal use only, achieving an 84% success rate in attacking GPT-5.1 in new scenarios, far exceeding the 13% of human red teams.
It successfully compromised a vending machine agent to lower prices and cancel orders, and induced Codex CLI to leak sensitive data.
Source: Public Information
ABAB AI Insight
OpenAI has been iterating on its security testing framework, and GPT-Red represents a leap from human red teaming to automated self-adversarial approaches, reflecting its strategy of balancing alignment and safety in the training of the GPT series.
The model utilizes self-play to mobilize internal computational resources to discover vulnerabilities, directly injecting red team results into the adversarial training of GPT-5.6 with the motivation to maximize defensive capabilities before deployment.
Similar to how early AlphaGo broke through Go boundaries through self-play, OpenAI is currently at the forefront of enhancing large model security, with automated red teaming becoming a core tool to address prompt injection and agent risks.
Essentially, this is a technological substitution, where AI-driven red team systems replace human experts for greater efficiency, achieving higher coverage and iteration speed, transforming models from being easily compromised to high robustness, and reconstructing AI security barriers.
ABAB News · Cognitive Laws
The stronger the attacker, the faster the defender evolves.
Self-play is the optimal path for security training.
The speed of vulnerability discovery determines the model's lifespan.