OpenAI today introduced GPT-6 Cyber, a specialist model built to help security teams discover, validate and remediate software vulnerabilities while operating under explicit authorization and safety limits. GPT-6 Cyber is intended to accelerate defensive research by combining greater offensive capability with enhanced alignment and tightened access controls, enabling researchers to probe targets in controlled settings and produce actionable reports for maintainers.
Benchmark results cited by OpenAI show clear gains over earlier cyber-focused models. On ExploitGym, which tests whether a model can exploit a specified vulnerability on a remote target, GPT-6 Cyber scored 48.5%, up from 32.1% for GPT-5.6 Cyber. In offensive web penetration testing—where models hunt for vulnerabilities in web applications without source access—GPT-6 Cyber achieved 40.3% recall of discovered vulnerabilities versus 32.0% for GPT-6 Astra. OpenAI also ran the model on a new RemoteExploitBench that contains known issues in applications such as GitLab, PostgreSQL and Jenkins in a controlled environment. There, GPT-6 Astra with Daybreak Blue refused all tasks, while GPT-6 Cyber completed 95% of attempts; the single failure involved an exploit that could not be triggered remotely. OpenAI elevated GPT-6 Cyber to a Critical rating under its Preparedness Framework, up from High for GPT-5.6 Cyber.
A major development in GPT-6 Cyber is its emphasis on limiting out-of-scope activity and improving the accuracy of reporting. OpenAI says the model underwent specialized training and evaluation designed to teach it to respect rules of engagement and the researcher’s authorization. In tests informed by past incidents, GPT-6 Cyber avoided exploiting deliberately easy-to-attack supporting services in 99.93% of trials when run without deployment safety prompts, compared with 94.49% for GPT-5.6 Cyber; with deployment prompts enabled, GPT-6 Cyber stayed within task bounds in 100% of tests. OpenAI reports these alignment improvements show up in the model’s internal reasoning, which more explicitly considers allowed targets and discourages probing credentials or services beyond the defined scope.
OpenAI also highlights early, real-world results. Researchers used GPT-6 Cyber to reproduce 1,724 vulnerabilities and produce working exploits for 948 Linux kernel issues, then disclosed 304 critical findings to maintainers along with candidate patches. In Mozilla’s Firefox codebase, the model supplied evidence and reproducible tests for 81 previously unknown high-severity vulnerabilities and indicated that 30 additional issues were more serious than initial estimates; Mozilla began landing fixes after receiving those reports. OpenAI further cites joint work that used GPT-6 Cyber alongside other models to develop and validate exploits showing remote code execution in downstream applications for a Brotli compression library flaw. The company says security teams at organizations including Cisco, Calif and CodeRabbit are using the model for vulnerability research and code review, and OpenAI is also applying it internally to harden the infrastructure used to train frontier models.
To reduce misuse risks, OpenAI paired GPT-6 Cyber’s capabilities with expanded safeguards and monitoring. Controls include cyber-specific misalignment monitoring that checks actions against stated authorization, targeted monitoring for emerging threats, automated Private Safety Processing available to some API customers, and default IP allowlisting for Daybreak Red API access. OpenAI recommends sandboxing, human review of actions and other oversight for high-risk workflows. The company also collaborated with Dragos to assess effects on operational technology and industrial control systems (OT/ICS) in laboratory settings; feedback from those evaluations helped refine protections aimed at preventing assistance that could cause unsafe device actions or service disruptions while preserving authorized testing capabilities.
Access to GPT-6 Cyber is initially limited. The model is available via API and Codex to a select set of approved Daybreak Red customers, with phased expansion planned. OpenAI is making GPT-5.6 Cyber available to all Daybreak Red customers and expanding Daybreak Blue alpha access for GPT-6 Astra to additional organizations and individual security researchers. The company also highlighted a $1 billion Daybreak for Frontline Defenders grant program intended to subsidize access for resource-constrained teams defending critical services.
OpenAI positions GPT-6 Cyber as a tool to broaden defensive capacity across maintainers and security teams while stressing layered safeguards and controlled distribution to mitigate misuse. Detailed deployment and evaluation information is available in OpenAI’s system card and associated reports, which the company says will guide continued refinement with partners and the broader security community.
Source: Read the original source
Leave a Reply