GPT-Red: Unlocking Self-Improvement for Robustness

What changed. OpenAI trained an automated red-teaming model through self-play and used it in production training; it reports sixfold fewer failures on its hardest direct prompt-injection benchmark versus its best production model four months earlier.

Technique / discovery. Self-play red teaming, attack generation, adversarial training, and held-out robustness evaluation.

Why it matters. Safety work is becoming an adversarial training flywheel in which models generate scalable attacks and new training data.

Application. Tool-using agents, browsers, and connected applications

Evidence maturity. Technical report (source tier A)

Identified bottleneck. Attackers and environments co-evolve; benchmark saturation can hide unmeasured failure modes.

Caveat / evidence note. Most evaluation infrastructure is internal and vendor reported.

Review status. Reviewed. User requested: Yes.

• OpenAI (opens in a new tab)

A reproduction tutorial is not yet available for this entry. The closest reviewed material is Safety, security, and alignment.