Background
A large language model capable of uncovering 81 previously unknown critical vulnerabilities in the long-scrutinized Firefox browser, while simultaneously developing functional exploit code for 948 vulnerabilities in the Linux kernel.
This is not a futurist's prediction of AI capabilities, but the track record of OpenAI’s newly released cybersecurity-focused model: GPT-6 Cyber.
OpenAI has officially launched GPT-6 Cyber, a model specifically engineered for cybersecurity research. Unlike its predecessors, which focused primarily on coding, reasoning, and general-purpose agency, this new model prioritizes vulnerability discovery, exploit validation, attack path analysis, and security remediation.
About the Data
According to OpenAI’s data, GPT-6 Cyber achieved a 48.5% success rate in the ExploitGym benchmark—a significant 51% relative improvement over its predecessor, GPT-5.6 Cyber (32.1%). In RemoteExploitBench, another assessment for remote exploitation, the new model successfully completed 95% of the test cases.
More importantly, GPT-6 Cyber is already being deployed in real-world software security research. Beyond Firefox and Linux, major organizations like Cisco, Palo Alto Networks, and CodeRabbit are utilizing it for vulnerability research and defensive security reviews. OpenAI has classified the cybersecurity capabilities of GPT-6 Cyber as "Critical" under its Preparedness Framework, a step up from the "High" rating assigned to GPT-5.6 Cyber.
1. Beyond GPT-6 Astra: Specialized in Vulnerability Mining and Exploit Validation
GPT-6 Cyber is not merely an incremental update to a general-purpose model. While the previously released GPT-6 Astra demonstrated impressive security capabilities—achieving 100% on ExploitBench—it was intentionally constrained to prevent potential misuse, such as refusing to generate proof-of-concept (PoC) exploit code.
GPT-6 Cyber is fine-tuned specifically for professional security research. OpenAI reports that it outperforms Astra on most cybersecurity tasks, even when Astra is not restricted. Notably, in web penetration testing, GPT-6 Cyber achieved a vulnerability recall rate of 40.3%, surpassing Astra’s 32.0%. This task requires the model to identify vulnerabilities through interaction with the target application without access to source code, simulating a real-world black-box testing environment.
2. Uncovering 81 Unknown Critical Firefox Vulnerabilities and 948 Linux Paths
If benchmarks represent the upper bound of capability, the findings in real-world open-source projects prove the model's practical value.
Firefox: Despite being one of the most thoroughly scrutinized open-source projects, Firefox was found by GPT-6 Cyber to contain 81 previously unknown "Critical" vulnerabilities. The model also provided evidence suggesting that another 30 known issues were more severe than previously assessed. Mozilla has already begun fixing these issues, with some patches having been deployed.
Linux Kernel: In a massive research effort detailed in the article "A thousand fewer paths to root," OpenAI built an automated pipeline for vulnerability research. The result: 2,308 candidate issues were identified, leading to 1,724 dynamically reproduced vulnerabilities and 948 valid local privilege escalation (LPE) exploits. This is a massive leap from the 166 exploits achieved by GPT-5.6 Sol under similar conditions. The cost for this research was approximately 1.21million,averaging1,279 per successful exploit—a cost-effective alternative to manual investigation.
3. AI Learning to Respect "Attack Boundaries"
A critical advancement in GPT-6 Cyber is its "Safety Alignment." For cybersecurity agents, raw power must be matched by strict adherence to scope.
OpenAI introduced the ExploitGym Honeypot test to measure whether the model respects task boundaries. When presented with a complex exploit task, the environment includes tempting, vulnerable auxiliary services. While the model could exploit these to achieve its goal faster, the correct behavior is to ignore them to remain within the authorized scope.
GPT-6 Cyber demonstrated a 99.93% compliance rate (avoiding unauthorized targets), compared to 94.49% for its predecessor. With deployment safety prompts enabled, this compliance rate reaches 100%. This proves that OpenAI is training the model not just how to find vulnerabilities, but how to judge what should not be done.
Access and Deployment Strategy
GPT-6 Cyber is not currently available to the general public. It is being rolled out to select "Daybreak Red" customers via API and Codex. OpenAI has tiered its security offerings:
Daybreak Blue: For verified defensive researchers, focusing on validation and remediation.
Daybreak Red: For authorized professionals, allowing deeper exploitation and testing. This tier includes mandatory IP whitelisting, specialized behavior monitoring, and strict access controls.
The release of GPT-6 Cyber reflects a key shift in the industry: the challenge is no longer just building a model that can find bugs, but creating a governance framework that empowers security researchers while preventing malicious actors from lowering their technical barrier to entry.