OpenAI, “GPTRed: Unlocking Self-Improvement for Robustness”
OpenAI introduced GPTRed, an internal automated red-teaming model trained through self-play to discover prompt-injection and agentic-system vulnerabilities and generate adversarial training data for production models. In an internal replication of a published prompt-injection challenge, GPTRed succeeded in 84% of novel scenarios versus 13% for human red-teamers; it also compromised a live autonomous vending agent by altering prices, ordering an expensive product at the minimum permitted price, and cancelling another customer’s order.
OpenAI introduced GPTRed, an internal automated red-teaming model trained through self-play to discover prompt-injection and agentic-system vulnerabilities and generate adversarial training data for production models.
Why it matters
OpenAI reports that adversarial training reduced GPT5.6 Sol’s failures sixfold relative to its strongest production model four months earlier and lowered direct GPTRed prompt-injection success to 0.05%. The results are company-reported and primarily based on internal evaluations, but they provide unusually concrete evidence that agentic systems remain vulnerable while also showing a potential safety-scaling mechanism.
Primary trail
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
