OpenAI has released GPT-6 Cyber, a specialized artificial intelligence model designed for authorized cybersecurity research that successfully identified and exploited 948 local privilege escalation vulnerabilities in the Linux kernel. The model achieved this during a single research run in mid-October 2026, demonstrating a significant increase in capability over previous versions.
GPT-6 Cyber scales vulnerability discovery and exploitation
During its testing phase, the new model utilized an autonomous pipeline to analyze 2,308 candidate findings within the Linux kernel. This process included dynamically reproducing 1,724 vulnerabilities and producing 948 working local privilege escalation (LPE) exploits, which could allow an unprivileged user to gain root access, OpenAI reported.
The research identified at least 10 novel Linux exploitation techniques that were previously undocumented, according to The Hacker News. These new methods include two new persistent elastic objects for spraying objects inside the target kernel cache, two new page grooming techniques to improve reliability, five new data-only techniques to obtain privilege escalation, and one technique involving arbitrary type confusion.
OpenAI’s findings, which were published as part of its Patch the Planet
initiative, highlight the model’s ability to navigate complex codebases. Scout agents first identified security-relevant sections of the Linux kernel, narrowing the audit to 5,034 targets covering approximately 2.8 million lines of code, or about 7 percent of the total codebase.
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline. In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API… pic.twitter.com/Nu2I0n2ayI
Comparison of model performance and safety boundaries
GPT-6 Cyber shows a marked improvement in both offensive capability and adherence to researcher-set constraints. In the ExploitGym benchmark, which measures an AI’s ability to exploit specific vulnerabilities in remote targets, GPT-6 Cyber scored 48.5 percent. This is approximately 16 percentage points higher than the 32.1 percent score achieved by GPT-5.6 Cyber, and a few points higher than the 42.4 percent score of GPT-6 Astra, The Verge reported.
Regarding safety and scope, the model demonstrated improved reliability in staying within authorized boundaries. Fouad Matin, the head of OpenAI’s cyber preparedness and safety team, stated during a press briefing that the model avoided exploiting out-of-scope services over 99 percent of the time. The Register noted that without deployment safety prompts, the model went out of bounds in 0.07 percent of cases, down from approximately 5 percent with GPT-5.6 Cyber.
It’s over 99 percent of the time doing what you ask,
Matin said. However, the issue of evaluation awareness—where an AI recognizes it is being tested and alters its behavior—remains a priority. Matin stated, Evaluation awareness is super top of mind for us.
API costs and hypothetical bounty valuations
The autonomous research run was resource-intensive, totaling $1,213,430 in API credits. Gizmodo reported that this averages out to $1,279 per LPE exploit.
OpenAI compared these costs to traditional human-led bug bounty programs. By applying historical bounty amounts of $20,000 to $70,000 to the 948 exploits, the company calculated a hypothetical valuation of $18.9 million to $66.3 million.
Impact on Firefox and Linux software maintenance
Beyond the Linux kernel, the model was used to identify 81 previously unknown, high-severity vulnerabilities in the Mozilla Firefox browser. Fortune reported that Mozilla has already begun landing fixes for many of these reported issues.
The full run cost $1,213,430 in API credits—about $1,279 per distinct local privilege escalation vulnerability. For context, applying historical Pwn2Own and TyphoonPWN bounty rates gives a hypothetical value of $18.9M–$66.3M. This demonstrates the relative efficacy of using AI…
Brian Grinstead, a Distinguished Engineer at Firefox, noted that even after multiple generations of frontier models provided hardening, the model uncovered unknown high-severity vulnerabilities and helped ship fixes to millions of users. In the Firefox campaign, the model also identified 30 previously reported issues.
The pipeline’s workflow includes a five-stage process: vulnerability discovery, dynamic validation, exploitability assessment, patch generation, and human review.
The pipeline finds security-relevant code, scans for vulnerabilities, reproduces them in isolated environments, tests whether they can be exploited, and generates candidate patches. Human researchers review the results before reporting vulnerabilities and propose fixes to… pic.twitter.com/xF5Scq69tm
Data from the Google Threat Intelligence Group (GTIG) shows the need for such automated tools. Insights from GTIG show that monthly disclosed vulnerabilities more than doubled between January 2026 (5,045) and August 2026 (10,740). High-risk disclosures also grew by 167% in that same period, rising from 131 in January to 350 in August.
Access through the Daybreak program
Due to its ability to identify and develop functional zero-day exploits in hardened real-world critical systems, OpenAI has classified GPT-6 Cyber as “Criticalunder its Preparedness Framework. The Verge reported that this is a higher risk classification than the
High” rating previously assigned to GPT-5.6 Cyber.
Access is currently restricted to select members of the Daybreak Red program, which is intended for organizations performing authorized defensive cybersecurity work. Marketscale noted that these customers can access the model through the API and Codex. OpenAI plans to expand access to all Daybreak Red customers in the coming weeks.
OpenAI is also expanding its Daybreak Blue alpha access, which provides GPT-6 Astra to more organizations and individual security researchers. The company is also creating a new access level specifically for critical infrastructure operators to perform authorized testing on their own systems.
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment. The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and… pic.twitter.com/34uAeNnhwP
Earlier coverage: OpenAI Pipeline Identifies 948 Linux Kernel Vulnerabilities
The run reproduced 1,724 vulnerabilities. It produced proofs of concept for 291 information leaks and working exploits for 948 distinct local privilege escalation vulnerabilities. Those exploits worked on a stock Linux desktop configuration with standard security protections…