OpenAI unveils automated red-teaming system GPT-Red

OpenAI on July 16 disclosed GPT-Red, an internal automated red-teaming system aimed at scaling security testing as model capabilities grow. The focus is prompt injection and similar weaknesses in tool-using AI agents, and the project matters because OpenAI frames manual red-teaming as a scaling bottleneck for safety and alignment work.

How it works

According to OpenAI, GPT-Red runs an adversarial self-play loop: one side continuously tries to break defender models with prompt-injection attacks, while successful attacks are converted into training data for the defensive side. The goal is to create a closed loop that keeps expanding coverage of failure modes, including broader and more complex attack patterns.

Reported results

OpenAI says GPT-5.6’s resistance to prompt injection improved by 6x after training with GPT-Red. In replay tests using GPT-Red’s strongest attacks that were not seen during training, OpenAI says GPT-5.6 Sol performed best and is currently its most robust model on prompt injection.

Reactions and significance

OpenAI researcher Kathy said that GPT-Red, as a model trained specifically to discover vulnerabilities, is already clearly outperforming human experts on red-team tasks. Across posts discussing the release, GPT-Red is portrayed less as a one-off benchmark result and more as security infrastructure: an automated AI-vs-AI sparring loop that can be plugged directly into model hardening and future training.

2026-07-16 ~ 2026-07-16 · 16 related posts

• GPT-Red: The Automated Red Teaming Model — btibor91 · 2026-07-16
• OpenAI's GPT-Red Safety Adversarial Training — nordicinst · 2026-07-16
• [source] OpenAI Launches GPT-Red Red Teaming System — OpenAI · 2026-07-16
• [source] GPT-Red Finds Vulnerabilities via Self-Play — OpenAI · 2026-07-16
• [source] GPT-5.6 Prompt Injection Resistance Jumps 6x — OpenAI · 2026-07-16
• GPT-Red: Automated AI Red-Teaming Model — yoimnotkesku · 2026-07-16
• GPT-5.6 Enhances Prompt Injection Defenses — Moh1tAgarwal · 2026-07-16
• GPT-Red Has Surpassed Human Red Team Levels — kaicathyc · 2026-07-16
• OpenAI Red Team Model GPT-Red Far Surpasses Humans — kaicathyc · 2026-07-16
• OpenAI's Internal Red Team Model GPT-Red Revealed — etherd0t · 2026-07-16
• GPT-Red Used to Train More Stable Models — Dr_Singularity · 2026-07-16
• OpenAI's Self-Play Red Team Flywheel — imjustnewatai · 2026-07-16

4 near-duplicate retellings: _AndrewZhao · shi_weiyan · CodeByPoonam · steipete