OpenAI launches GPT-6 Cyber: New AI cybersecurity model finds major software flaws

OpenAI has launched GPT-6 Cyber, a new cybersecurity model engineered to identify software security weaknesses, develop methods to test them, and assist security teams in patching vulnerabilities.

The company stated the model identified 81 previously unknown high-severity vulnerabilities in the Firefox browser and constructed working exploits for 948 vulnerabilities in the Linux kernel, underscoring its capacity to detect and evaluate software security risks.

OpenAI said, “We’re introducing GPT-6 Cyber, a major leap in security research: more capable at finding and fixing vulnerabilities, and far more reliable staying within the boundaries researchers set”.

According to OpenAI, Mozilla categorised the 81 Firefox vulnerabilities as high severity. OpenAI provided evidence, reproducible test cases, and proposed patches to Mozilla, which began implementing fixes shortly after receiving the disclosures. The model also flagged 30 known Firefox issues that were more severe than previously documented.

In its Linux kernel evaluations, GPT-6 Cyber successfully reproduced 1,724 vulnerabilities and developed working exploits for 948 of them.

OpenAI noted this represented more than five times the 166 vulnerabilities exploited by its earlier GPT-5.6 Sol model utilising the identical testing environment. The firm also disclosed 304 critical findings to maintainers alongside candidate patches, while remediation efforts for other issues remained underway.

The model registered a score of 48.5 per cent on the ExploitGym benchmark, in contrast to 32.1 per cent achieved by GPT-5.6 Cyber. OpenAI characterised this as a 51 per cent relative performance gain over the predecessor model.

On a newly established web penetration testing benchmark, GPT-6 Cyber achieved a recall score of 40.3 per cent, compared to 32 per cent for GPT-6 Astra.

In another evaluation, RemoteExploitBench, GPT-6 Cyber executed 95 per cent of assigned tasks, whereas GPT-6 Astra equipped with Daybreak Blue declined all tasks. OpenAI stated the sole task GPT-6 Cyber failed to complete involved an exploit it was unable to execute remotely.

OpenAI also emphasised enhancements in constraining the model within authorised testing parameters. It refrained from exploiting services outside the approved scope in 99.93 per cent of tests conducted without a deployment prompt, reaching 100 per cent when the prompt was applied. Its predecessor, GPT-5.6 Cyber, recorded 94.49 per cent compliance without prompts.

The organisation has designated GPT-6 Cyber as "Critical" for cybersecurity capabilities under its internal Preparedness Framework, upgrading from the "High" rating assigned to GPT-5.6 Cyber.

However, the advanced capabilities of the model have raised apprehension regarding prospective misuse.

Dragos, a global provider of extended operational technology (xOT) cybersecurity, conducted evaluations using cyber-permissive Daybreak Red models within a controlled laboratory setting. During tests executed in a simulated power plant facility, GPT-6 Cyber uncovered two previously unknown vulnerabilities in engineering software utilised for managing protective relay data.

These flaws could potentially expose protection settings and circuit-breaker trip logic. OpenAI stated it reinforced the model's safety guardrails following the trial findings.

The company indicated it has instituted cyber-specific monitoring, threat tracking, sandboxing, human oversight, and supplementary safeguards tailored for critical infrastructure environments.

OpenAI is rolling out GPT-6 Cyber initially via its API and Codex to select approved Daybreak Red clients, with access planned to expand incrementally over time. Daybreak Red is tailored for advanced exploit development and authorised red-teaming, whereas Daybreak Blue concentrates on validating vulnerabilities and engineering remediations.

Additionally, the firm announced a USD 1 billion Daybreak grant initiative aimed at frontline cybersecurity defenders and resource-constrained organisations, including government entities and essential service providers. It has also established a dedicated subgroup allowing verified critical infrastructure defenders to conduct authorised security testing.

ANI