OpenAI has unveiled a new artificial intelligence system named GPT-Red. This innovative model is aimed at proactively identifying security vulnerabilities within OpenAI's own AI frameworks before they can be exploited by malicious actors.
What is GPT-Red?
GPT-Red is designed as a red teaming tool, a process typically performed by cybersecurity professionals to assess the weaknesses in systems. Unlike traditional methods, GPT-Red automates this testing using AI, enabling it to conduct a much larger volume of probing attacks than human testers could manage.
How GPT-Red Works
The operational mechanism of GPT-Red involves sending a variety of prompts to targeted AI models and analyzing their responses. In cases where an attack fails to reveal a vulnerability, GPT-Red adapts its approach and continues testing until it achieves the desired outcome.
Training Methodology
OpenAI explains that GPT-Red is trained using a technique called reinforcement learning through self-play. This involves the attacker AI engaging with a defender AI in multiple scenarios. The attacking model receives rewards for successfully exploiting a system, while the defending model is rewarded for thwarting attacks. This iterative process allows both models to gradually enhance their capabilities.
Results of Internal Testing
According to OpenAI, GPT-Red has demonstrated a significantly higher success rate in identifying security flaws compared to human security teams. In internal tests, GPT-Red successfully exploited 84% of the scenarios it was presented with, whereas human testers managed only a 13% success rate.
Moreover, the introduction of GPT-Red has reportedly reduced the effectiveness of prompt injection attacks compared to models released just four months ago. Notably, a type of attack known as fake chain-of-thought, which previously had a success rate exceeding 95% against the GPT-5.1 model, now has dropped to below 10% against the newer GPT-5.6.
Testing Against AI Agents
GPT-Red was also tested against various AI agents. In one notable experiment, it managed to take control of a virtual vending machine agent called Vendy, altering product prices and even canceling orders. Additionally, it was able to manipulate command-line-based AI agents that assist in programming tasks.
“Compared to human red teaming teams, this model is exceptionally effective at uncovering successful attack vectors,” stated OpenAI researcher Dylan Hunn.
Source: https://tekno.kompas.com/read/2026/07/16/17040067/openai-kenalkan-gpt-red-senjata-untuk-melawan-ai-nya-sendiri
