OpenAI has introduced GPT‑6 Cyber, a model designed specifically for cybersecurity research. According to the company, the model improves both exploit discovery and alignment with researcher‑specified constraints compared with GPT‑5.6 Cyber and outperforms GPT‑6 Astra on many cyber tasks. OpenAI reports benchmark gains, large‑scale findings in real software, and a suite of safeguards and access controls for sensitive use.
Performance and benchmarks
- ExploitGym: GPT‑6 Cyber scored 48.5% on ExploitGym, which evaluates whether a model can exploit a specified vulnerability on a remote target; GPT‑5.6 Cyber scored 32.1%. This represents roughly a 51% relative improvement.
- Scope adherence: In an alignment evaluation measuring whether the model avoids exploiting out‑of‑scope services, GPT‑6 Cyber avoided such exploitation in 99.93% of tests without a deployment prompt and 100% with the deployment prompt.
- Offensive web testing: In a new evaluation for offensive web penetration testing (probing web applications without source code access), GPT‑6 Cyber achieved 40.3% recall of vulnerabilities versus 32.0% for GPT‑6 Astra.
- RemoteExploitBench: In a benchmark that deploys known vulnerabilities in applications such as GitLab, PostgreSQL, and Jenkins in a controlled environment, GPT‑6 Cyber completed 95% of tasks in an initial evaluation; GPT‑6 Astra running with Daybreak Blue refused 100% of those tasks. GPT‑6 Cyber’s sole failure was an exploit it could not get to work remotely. Full evaluation settings and results are described in the system card.
Real‑world findings and use
OpenAI reports several concrete results where GPT‑6 Cyber assisted vulnerability research and remediation:
- Linux kernel: researchers reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. In line with Linux community guidance, 304 critical findings were disclosed to maintainers with candidate patches.
- Firefox: during a research campaign GPT‑6 Cyber produced reproducible tests and evidence for 81 previously unknown vulnerabilities classified by Mozilla as high severity, and identified 30 additional known issues as more severe than previously reported. The findings and proposed patches were disclosed to Mozilla, which began shipping fixes shortly after receiving the reports.
- Brotli RCE: researchers used GPT‑6 Cyber along with other OpenAI models to develop and validate exploits for a critical Brotli vulnerability that could enable remote code execution in downstream applications.
OpenAI also used GPT‑6 Cyber to help harden its own infrastructure. In controlled red‑teaming exercises, agents using GPT‑6 Cyber and other models chained vulnerabilities to escape microVMs and execute code on underlying hosts; the company used these findings to patch software, tighten network restrictions, and report issues to maintainers and vendors.
Access, Daybreak programs, and user restrictions
- Availability: GPT‑6 Cyber is initially available via the API and Codex to a limited set of approved Daybreak Red customers; access will be expanded progressively to all Daybreak Red customers. GPT‑5.6 Cyber will be made available to all Daybreak Red customers. Daybreak Blue alpha access for GPT‑6 Astra is also being expanded to more organizations and individual security researchers.
- Daybreak programs: the Daybreak for Frontline Defenders grant program ($1 billion) subsidizes access for resource‑constrained teams such as government organizations and essential service providers. A new group, Daybreak Red for Critical Infrastructure, will support approved critical infrastructure defenders with permissions for authorized testing while retaining safeguards.
Safeguards and oversight
OpenAI paired GPT‑6 Cyber’s capabilities with multiple protective measures:
- Cyber‑specific misalignment monitoring: monitoring is aimed at detecting when actions diverge from a user’s authorization while allowing legitimate exploit development and testing.
- Targeted monitoring for emerging threats: the company established focused monitoring during incidents such as coordinated disclosures of critical PHP vulnerabilities to detect exploit development attempts against live servers while preserving defender support.
- Private Safety Processing (PSP): automated checks review related interactions for potential misuse. API customers using Zero Data Retention typically must use PSP, which reviews encrypted records retained in customer storage for up to 30 days without exposing prompts and responses to OpenAI staff.
- Credential protections: Daybreak Red API access requires IP allowlisting by default and monitoring for credential compromise.
- Sandboxing and human oversight: OpenAI recommends sandboxing, action review, and human oversight for high‑risk workflows.
- Protections for critical systems: safeguards were refined to block assistance and capability development that could compromise industrial control systems or safety‑critical equipment, interfere with operations, or cause serious physical harm.
Industry evaluations and partnerships
OpenAI collaborated with Dragos, a provider of extended operational technology (xOT) cybersecurity, to evaluate Daybreak Red models in a controlled lab environment. Dragos’s tests covered workflows from network discovery and asset enumeration to targeted reconnaissance, sensitive data exfiltration, and potentially dangerous device actions. Dragos feedback was used to refine safeguards to block harmful assistance while minimizing impact on authorized testing. OpenAI is pursuing further partnerships with offensive security providers, purple teams, energy companies, critical infrastructure operators, and national labs to continue assessing models and safeguards.
Risks and mitigation
OpenAI acknowledges increased misuse risk as model capabilities grow and reports pairing stronger capabilities with layered safeguards to detect and disrupt misuse and keep model behavior within authorized intent. The company continues to solicit feedback from partners and the security community about gaps or interruptions in safeguards.
Conclusion
GPT‑6 Cyber represents a step change in OpenAI’s offerings for security research: measurable benchmark improvements, reported large‑scale findings in major codebases, and expanded controls to govern sensitive use. The model is initially available to a limited set of Daybreak Red customers, with plans to expand access and support through the Daybreak programs for frontline defenders, critical infrastructure teams, independent researchers, and open‑source maintainers.
(Details, system card results, and evaluation settings are described in OpenAI’s documentation and system card.)
