Openai Unveils Gpt-red: Automated Model Outsmarts Human Testers

WHO:

MarkTechPost - OpenAI is the organization that developed GPT-Red.

WHAT:

MarkTechPost - GPT-Red is an internal automated red-teaming model that outperformed human red-teamers with a score of 84% to 13% on prompt injection tests.

MarkTechPost - GPT-Red discovered a new attack class termed "Fake Chain-of-Thought."

MarkTechPost - The model reduced failures of GPT-5.6 Sol by six times on OpenAI's most challenging direct injection benchmark.

Techmeme - GPT-Red is designed to find and fix prompt injection vulnerabilities at scale before wider deployment.

WHERE:

MarkTechPost - The testing for GPT-Red took place in a replicated indirect prompt injection arena.

WHEN:

Techmeme - OpenAI aims to deploy GPT-Red before wider use of its models.

WHY:

Techmeme - The development of GPT-Red is part of OpenAI's efforts to improve the robustness of its models through strong automated safety red-teamers.

HOW:

MarkTechPost - GPT-Red was trained using self-play reinforcement learning against a population of defender LLMs.

MarkTechPost - OpenAI acknowledges that GPT-Red still faces challenges with multi-turn and image-based attacks.

(UPDATED July 16, 2026 19:24 UTC)

OpenAI has developed a new internal automated red-teaming model called GPT-Red, which achieved an impressive score of eighty-four percent to thirteen percent against human red-teamers in prompt injection tests. This model not only discovered a new attack class known as "Fake Chain-of-Thought," but it also significantly reduced failures of GPT-5.6 Sol by six times on OpenAI's toughest direct injection benchmark. OpenAI plans to deploy GPT-Red before the wider use of its models to enhance their robustness and address prompt injection vulnerabilities at scale.