Overview of GPT-6 Cyber
OpenAI has unveiled GPT-6 Cyber, a specialized artificial intelligence security model designed to assist with automated vulnerability discovery and authorized penetration testing. According to SecurityWeek, the new model demonstrates a substantial leap in capability compared to its predecessor, scoring 48.5% on the ExploitGym benchmark—an evaluation framework testing whether an AI can successfully exploit remote flaws—compared to 32.1% for the older GPT-5.6 Cyber variant.
Despite classifying the model as Critical for cybersecurity under its internal Preparedness Framework, OpenAI highlighted significantly improved alignment metrics. During controlled tests involving intentional decoys and restricted targets, the model avoided unauthorized exploitation attempts in 99.93% of unprompted evaluations, and 100% of cases when deployment safety prompts were active.
Technical Capabilities and Findings
During rigorous internal evaluations, researchers pointed the model at complex, widely deployed software targets to measure its efficacy in vulnerability research.
- Linux Kernel: The model successfully reproduced 1,724 vulnerabilities and constructed functional privilege escalation exploits for 948 of them, subsequently reporting 304 critical issues along with proposed fixes to maintainers.
- Mozilla Firefox: The system documented 81 previously unknown flaws categorized as high severity by Mozilla, supplying reproducible tests that allowed developers to begin rolling out patches.
Refinements to safety guardrails were shaped by collaboration with industrial cybersecurity firm Dragos, whose experts tested early iterations in a controlled laboratory environment. The model access remains tightly restricted, currently available only to select customers through the API and Codex within the Daybreak Red program designed for authorized security testing.
Defensive Implications and Mitigations
While advanced models assist researchers in finding and remediating flaws before malicious actors weaponize them, the lowered skill floor for reconnaissance presents ongoing strategic challenges for enterprise defenders.
- Continuous Patching: Ensure automated ingestion pipelines for vendor updates are active, particularly as AI-driven discovery accelerates the public disclosure of zero-day and edge-case vulnerabilities.
- Scope Enforcement: Security teams deploying agentic frameworks or AI assistants must implement strict guardrails and runtime monitoring to verify that automated actions align with authorized boundaries.
- Threat Exposure Validation: Incorporate automated red-teaming scenarios into continuous threat exposure management programs to evaluate enterprise resilience against rapid, machine-speed reconnaissance.
Related: OpenAI GPT-6 Cyber AI Finds Firefox Flaws and Linux Bugs, Microsoft Warns Threat Actors Lead the Early AI Security Race