
In short
- OpenAI launched GPT-Red, an automated AI system designed to find problems in GPT formats before they are released.
- The company said GPT-Red was used to train GPT-5.6, reducing failures on one of its most difficult injection benchmarks.
- The system is designed to support human red tape, third-party testing, and other AI security measures.
OpenAI has launched GPT-Red, an automated AI system designed to detect security vulnerabilities in its languages.
GPT-Red takes its name from cybersecurity red teaming, which is the practice of deliberately trying to break into a system to identify vulnerabilities before attackers can exploit them.
In a post on Wednesday, OpenAI said the tool helped create it GPT-5.6 It is highly resistant to injection immediately before delivery.
“As modeling capabilities grow, security and compatibility must grow,” OpenAI he wrote on X. “Combining red groups is important, but today’s methods are difficult to scale, which makes it a big problem. GPT-Red is one way we’re tackling it.”
According to OpenAI, GPT-Red he trained by learning to brace himself, allowing him to shoot faster and slower as defenders learned to resist. The company said that these displays were included in the GPT-5.6 training, stating that GPT-Red performed successfully in 84% of indoor lighting situations, compared to 13% of human red players in the same tests.
“GPT-Red learns through the adversary’s gameplay, where the goal is to introduce different types of defenses,” wrote OpenAI. “Every attack that GPT-Red finds is used to improve these defenders, pushing GPT-Red to continue to find bigger and more complex vulnerabilities.”
In a separate investigation, OpenAI said the system manipulated an independent system vendor into lowering prices, ordering discounted products, and canceling a customer’s order before disclosure and control.
GPT-Red follows years of cybersecurity and OpenAI efforts following the public launch of ChatGPT.
In 2023, the company he started Its OpenAI Red Teaming Network, employing external cybersecurity researchers and local experts to investigate ChatGPT and other types of security flaws before they are released. GPT-Red expands this effort by making more of the process, using an AI-based model quick injection attacks and other countermeasures at a level that would be difficult for human researchers alone.
The announcement of OpenAI marks a major shift in the use of AI to protect AI.
Earlier this month, the Ethereum Foundation said it did to be sent AI contributors to the red groups of the complex social network, revealing the vulnerability of software used by Ethereum consensus clients. Researchers said AI agents can search larger codebases than humans, but the problem has shifted from finding bugs to identifying which ones can be exploited.
According to OpenAI, GPT-Red will remain an internal tool because it contains intentionally offensive tools.
“We believe with GPT-Red that we have begun to release a common thread for security, where today’s models can be used to make tomorrow’s models stronger, more compatible, and more reliable,” he said.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.





