OpenAI has built an automated “red team” to hunt for the kinds of prompt-injection attacks that can compromise language models — and it says the system helped harden its newest model, GPT-5.6, before release.
Called GPT-Red, the tool mimics the cybersecurity practice of red teaming by intentionally trying to break models so weaknesses can be fixed before attackers exploit them. OpenAI says GPT-Red was trained with self-play reinforcement learning: it generates increasingly sophisticated prompt-injection attacks while “defender” models learn to resist them. Successful attacks discovered by GPT-Red are fed back into model training, creating a continuous adversarial loop that pushes both the attackers and defenders to improve.
OpenAI reported striking results in internal evaluations: GPT-Red succeeded in 84% of test scenarios, while human red teamers only succeeded in 13% of the same tests. The company also shared an illustrative case in which GPT-Red compromised an autonomous vending-machine agent, tricking it into lowering prices, ordering discounted stock, and canceling another customer’s order — vulnerabilities that were fixed once exposed.
GPT-Red builds on years of OpenAI’s security work. In 2023 the company launched the OpenAI Red Teaming Network to enlist external researchers and domain experts to probe ChatGPT and other models. GPT-Red automates much of that process, enabling red-team scale and speed that would be difficult for humans to match.
The move is part of a broader industry trend toward using AI to secure AI. Earlier this month the Ethereum Foundation said it had deployed AI agents to red-team critical network infrastructure and uncovered a vulnerability in software used by consensus clients. Researchers note AI agents can scan larger codebases and mount more varied tests than humans, though distinguishing theoretical bugs from truly exploitable issues remains a challenge.
OpenAI says GPT-Red will remain an internal tool because it intentionally contains offensive capabilities. The company framed the project as building a safety “flywheel” — using today’s models to make tomorrow’s models more robust, aligned, and trustworthy.
Why this matters to crypto: as decentralized finance, trading bots, oracle systems, and autonomous agents take on more responsibility in crypto ecosystems, prompt-injection and adversarial attacks on models embedded in those systems could create new vectors for manipulation. Automated red-teaming like GPT-Red could help model providers and protocol teams uncover and patch vulnerabilities at scale — but it also raises questions about disclosure practices, who gets access to internal findings, and how the industry validates that discovered flaws are truly mitigated.
Read more AI-generated news on: undefined/news
