What happened
OpenAI released GPT-6 Cyber, a new AI model specifically designed for cybersecurity research. The model demonstrates substantial improvements over GPT-5.6 Cyber, including a 51% relative improvement on the ExploitGym benchmark and enhanced alignment to stay within authorized scopes. It has already been used to identify 81 novel high-severity vulnerabilities in Firefox and develop exploits for 948 Linux kernel vulnerabilities. Access is currently limited to a select group of approved Daybreak Red customers via API and Codex, with broader expansion planned.
OpenAI has introduced GPT-6 Cyber, a specialized AI model designed to enhance cybersecurity research capabilities. The model shows significant performance gains over its predecessor, GPT-5.6 Cyber, particularly in identifying and exploiting software vulnerabilities. On the ExploitGym benchmark, which tests a model's ability to exploit specified vulnerabilities, GPT-6 Cyber achieved a score of 48.5%, compared to 32.1% for GPT-5.6 Cyber. This represents a 51% relative improvement, indicating a substantial leap in the model's offensive security capabilities.
The model has already been applied to real-world security challenges. OpenAI researchers used GPT-6 Cyber to identify 81 previously unknown high-severity vulnerabilities in Firefox, for which Mozilla has begun releasing fixes. Additionally, the model developed working exploits for 948 Linux kernel vulnerabilities, a figure more than five times higher than the 166 exploits developed by GPT-5.6 Sol using the same research harness. These findings have been disclosed to maintainers, with candidate patches provided to facilitate the fixing process.
A key focus of GPT-6 Cyber's development is alignment, ensuring the model stays within the boundaries set by researchers. In alignment evaluations, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without deployment prompts, and 100% with them. This improvement is attributed to dedicated training to respect authorized scopes and report results accurately. The model's chain of thought reveals its reasoning process in maintaining these boundaries, providing transparency into its decision-making.
Access to GPT-6 Cyber is currently restricted to a select group of approved Daybreak Red customers via the API and Codex. OpenAI plans to progressively expand access to all Daybreak Red customers and is also making GPT-5.6 Cyber available to a broader audience. The company has implemented several safeguards to prevent misuse, including cyber-specific misalignment monitoring, targeted monitoring for emerging threats, and Private Safety Processing (PSP) for API customers using . IP allowlisting is required for Daybreak Red API access by default, and sandboxing and human oversight are recommended for high-risk workflows.
Why it matters
The release of GPT-6 Cyber marks a significant escalation in the capability of AI systems to perform offensive security tasks, such as developing working exploits and identifying zero-day vulnerabilities in critical software like Linux and Firefox. This shift has profound implications for the cybersecurity landscape, as the model's ability to autonomously find and fix vulnerabilities could accelerate the patching process for critical infrastructure and open-source software. However, the increased capability also raises concerns about misuse by malicious actors, prompting OpenAI to implement stricter safeguards, including specialized monitoring and access restrictions. The model's as 'Critical' under OpenAI's Preparedness Framework underscores the potential impact of such tools on both defensive and offensive security operations, necessitating careful management of access and oversight.
The release of GPT-6 Cyber represents a significant advancement in the use of AI for cybersecurity, with the potential to dramatically improve the speed and efficiency of vulnerability discovery and patching. The model's ability to identify and exploit vulnerabilities in critical software like Linux and Firefox could help secure essential systems and reduce the risk of cyberattacks. However, the increased capability also raises concerns about the potential for misuse by malicious actors, who could use the model to develop more sophisticated attacks.
OpenAI's of GPT-6 Cyber as 'Critical' under its Preparedness Framework highlights the potential impact of such tools on the cybersecurity landscape. The company has implemented a range of safeguards to mitigate the risks of misuse, including specialized monitoring, access restrictions, and collaboration with security partners like Dragos. These measures aim to balance the need for powerful security tools with the imperative to prevent their misuse.
The model's development and deployment also have implications for the broader AI industry, as it demonstrates the potential for specialized AI models to address specific, high-stakes challenges. The focus on alignment and safety in the model's design reflects a growing recognition of the need for responsible AI development, particularly in domains where the consequences of misuse can be severe. The release of GPT-6 Cyber is likely to spur further innovation in AI-driven cybersecurity, as well as increased scrutiny of the safety and security measures in place to manage such powerful tools.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
Monitor the expansion of GPT-6 Cyber access to all Daybreak Red customers and the rollout of GPT-5.6 Cyber to a wider audience. Observe how security teams and organizations integrate the model into their vulnerability research and patching workflows. Watch for any reported incidents of misuse or attempts to distill the model by adversarial groups. Additionally, track the development of new benchmarks and evaluations, such as RemoteExploitBench, that assess the model's capabilities in controlled environments. The response of the broader security community and regulatory bodies to the release of such a powerful tool will also be significant.
The expansion of GPT-6 Cyber access to all Daybreak Red customers and the rollout of GPT-5.6 Cyber to a wider audience will be key developments to monitor. The impact of these models on the speed and effectiveness of vulnerability discovery and patching in real-world environments will be a critical factor in assessing their value and potential risks.
The response of the security community and regulatory bodies to the release of GPT-6 Cyber will also be significant. The model's capabilities and the safeguards in place to manage its use will likely be subject to scrutiny, and any reported incidents of misuse or attempts to distill the model by adversarial groups will be closely watched. The development of new benchmarks and evaluations, such as RemoteExploitBench, will provide further insight into the model's capabilities and the effectiveness of the safeguards in place.
The collaboration between OpenAI and security partners like Dragos, as well as energy companies and national labs, will be important in assessing the model's effectiveness and safeguards in industrial defense. The feedback from these partnerships will inform the refinement of the model's safeguards and the development of best practices for its use in critical infrastructure. The ongoing investment in threat investigations to detect model usage involving OT/ICS will also be a key area to monitor.