OpenAI Releases GPT-6 Cyber Under Restricted Access; Partners With Hacktron

OpenAI has introduced GPT-6 Cyber, a specialized variant of its GPT-6 Astra model designed specifically for authorized vulnerability discovery, exploit analysis, and network penetration testing. The artificial intelligence research organization classified the cybersecurity model as Critical under its internal Preparedness Framework, establishing strict operational containment protocols. Access to the offensive and defensive security model is strictly restricted through OpenAI's Daybreak Red program to vetted defensive security institutions. Participating organizations must satisfy rigorous identity verification requirements, including single sign-on integration, multi-factor authentication, and static API IP address allow-listing.

Hacktron Collaboration and Daybreak Red Program

Concurrently, OpenAI entered into a strategic collaboration with security firm Hacktron to discover and remediate security vulnerabilities across internal AI infrastructure. Hacktron researchers build autonomous AI agents designed for automated code inspection, red teaming simulations, and cloud service penetration evaluations. Through the partnership, OpenAI is granting Hacktron expanded access to its frontier cybersecurity models through Daybreak Red to evaluate third-party integrations. Hacktron Chief Executive Officer Zayne Zhang stated that combining offensive security expertise with frontier models helps protect the broader software ecosystem.

ExploitGym Benchmarks and Linux Codebase Testing

Benchmark evaluations reveal that GPT-6 Cyber scored 48.5% on the ExploitGym evaluation suite, outperforming GPT-5.6 Cyber's 32.1% result for a 51% relative performance gain. Safety audits demonstrated that the model avoided targeting out-of-scope digital infrastructure in 99.93% of evaluation runs without specific prompting, and maintained perfect boundary compliance when deployment safeguards were applied. The model was deployed across security-critical software repositories, successfully analyzing vulnerabilities in open-source foundations including the Linux kernel and the Firefox web browser. An experimental pipeline run reproduced 1,724 vulnerabilities within approximately 30 hours, generating functional proofs of concept for local privilege escalation and information disclosure flaws. Testing against the kernelCTF environment produced 25 information leak proofs of concept, according to OpenAI. OpenAI disclosed 304 critical Linux kernel vulnerability findings directly to open-source maintainers alongside proposed candidate patches, according to The Economic Times. In late July 2026, OpenAI's earlier cyber pipeline powered by GPT-5.6 Sol found and exploited 166 unique local privilege escalation vulnerabilities in the Linux kernel for $513,690 in API credits.

Pipeline Verification and Defensive Deployment Controls

These experimental outputs are pipeline candidates and automated proofs of concept, not patches already accepted by Linux maintainers. On the ProtocolQA evaluation benchmark, the model registered a standard score of 47.2%, increasing to 50.9% when safe refusals are tabulated. The rollout reflects rising demand from enterprise financial institutions and technology providers seeking structured, auditable frontier AI defenses against automated threat campaigns. OpenAI engineers and Hacktron researchers will choose high-priority operational areas to evaluate together, emphasizing security interfaces across external cloud platforms. Earlier in the week, the company stated that Cathedral researchers will collaborate alongside internal teams during an upcoming defensive security sprint. OpenAI also increased its bug bounty compensation to incentivize outside white-hat analysts to investigate potential system vulnerabilities. Within 10 days of completing OpenAI's autonomous scan in mid-October 2026, Linux maintainers merged upstream patches for 61 LPEs, prepared staging patches for 43, and publicly discussed 299 findings.