Skip to content

Back to the catalog

openai.comarticle

AI Security Briefing

OpenAI describes GPT-Red, an internal automated red-teaming model trained with self-play reinforcement learning against diverse defender LLMs to find prompt injections. It was used to adversarially train GPT-5.6 Sol, which fails on only 0.05% of GPT-Red's direct injections. GPT-Red succeeded on 84% of held-out indirect-injection arena scenarios versus 13% for humans, and it compromised a production AI vending-machine agent from Andon Labs. It also found a "Fake Chain-of-Thought" attack class that reached over 95% success on GPT-5.1 and now falls below 10%.

TL;DR OpenAI's GPT-Red is a self-play RL red-teaming model that generates prompt injections to adversarially train GPT-5.6 Sol, and it beat human red-teamers in held-out tests.
Open original ↗

Organisations named here

  • Andon Labs
  • Fake Chain-of-Thought
  • GPT-5.6 Sol
  • Project Vend

Technologies named here

  • GPT-5.1
  • GPT-5.5 and earlier OpenAI internal and production models
  • Project Vend
  • Codex CLI agent based on GPT-5.4 mini
  • OpenAI
  • GPT-5.5
  • Codex CLI
  • LLM
  • GPT-Red

Techniques named here

  • "Fake Chain-of-Thought" attacks
  • Data exfiltration via agent prompt injection
  • Automated red-teaming with self-play reinforcement learning
  • Scenario-based threat modeling
  • Prompt Injection

Also named here

  • Dziemian et al.

Related in AI Security