GPT-6 Cyber: 81 Firefox bugs and 948 Linux exploits, what the data shows

GPT-6 Cyber is an OpenAI model unveiled on November 5, 2026 to find vulnerabilities, reproduce them and propose fixes. According to the company, it found 81 previously unknown serious flaws in Firefox and produced 948 Linux exploits in the lab. It matters for businesses because part of researchers’ work can be organized into automated campaigns, but access is limited and controls are needed.¹²³

Who can use GPT-6 Cyber

The model is not public. For now it is available through the API and Codex to a select group of customers of Daybreak Red, the OpenAI program that covers advanced exploit development and authorized security testing.¹⁴³ OpenAI says it wants to extend access to all Daybreak Red customers soon. A parallel program, Daybreak Blue, lets verified defenders investigate flaws and validate fixes with fewer restrictions.³ In the Preparedness Framework, the framework OpenAI uses to measure the risks of its own models, GPT-6 Cyber is rated “Critical” for cyber capabilities, while the previous model was rated “High”. The label describes the level of capability, including the potential for offensive use, and not the reliability of the product.¹²³

The 948 Linux exploits: how they came about and what they cost

The data comes from an OpenAI study on the Linux kernel, published on November 4. In the October session, the agents started from 2,308 possible issues, confirmed 1,724 of them at runtime and built 948 local privilege escalation (LPE) exploits, that is, code that lets a user with limited access obtain administrator privileges.¹²³ The exploits show that problems already identified can be exploited in a test environment. They do not correspond to 948 observed attacks or to 948 new flaws. OpenAI forwarded 304 critical reports with patch proposals to maintainers, and the final review is up to the researchers.¹² The campaign cost $1,213,430 in API credits, about $1,279 per exploit. That is the cost of that experiment, not a price list for GPT-6 Cyber. The agents also limited the initial work to about 2.8 million lines of code, roughly 7% of the kernel.¹ OpenAI did not say which Linux distribution was tested. According to the site t3n, which cites several media outlets, it was reportedly a very widely used desktop system in its standard configuration.⁵

Firefox: 81 flaws and Mozilla’s fixes

In Firefox, OpenAI reports 81 new vulnerabilities, rated high severity by Mozilla, with reproducible proofs and candidate fixes. Another 30 already known issues turned out to be more serious than originally reported.¹² Mozilla has already distributed many of the patches. Administrators can check the security advisories, including those for Firefox 159, to find out which versions to update.¹²³ Among the results OpenAI cites is also a critical vulnerability in Brotli, a data compression component, with the identifier CVE-2026-111497 and a CVSS score of 9.2.⁴

Benchmarks and external tests: where the model falls short

On ExploitGym, a test in which the agent must turn known vulnerabilities into working exploits, GPT-6 Cyber succeeds in 48.5% of cases versus 32.1% for GPT-5.6 Cyber, 16.4 percentage points more. The figure comes from OpenAI and refers to a controlled environment.¹⁴³ The system card, the accompanying technical document, includes a check by Irregular, an outside firm. In CyScenarioBench the model managed to solve all 11 challenges in at least one attempt, with an average success rate of 61%. In FrontierCyber, on real software and hardware, it solved 90 problems out of 288, compared with 96 for GPT-6 Astra.² Neither model got past the 15 hardest tests, and Irregular did not observe successful attacks against fully hardened targets.² On staying within the authorized perimeter, OpenAI says GPT-6 Cyber did not exploit out-of-scope services in 99.93% of cases without additional safety instructions, versus 94.49% for GPT-5.6 Cyber, and in 100% with those instructions. In another evaluation, on reaction to warnings, it stopped at the right moment in 73% of trials, a share below the 78.38% achieved by GPT-6 Astra.²⁴

The OT front with Dragos

OpenAI worked with Dragos, a security company for OT (operational technology, the technologies that control plants and physical processes) systems, to test advanced models in laboratories that imitate real infrastructure. In the simulation of an energy plant, one model discovered two zero-day flaws, that is, flaws not yet known to vendors, in a program that handles protective relay information, including the settings and rules by which breakers trip.² Cyber Security 360 links these tests to the risk of operational disruption when an agent gets close to industrial devices.²

What changes for businesses

The results on Linux and Firefox indicate that the advantage lies in moving from simple reporting to proof of exploitability and a proposed patch. AI4Business suggests starting with known applications and measuring confirmed issues, false positives, regressions, time and costs.¹² The strongest numbers, however, are reported by OpenAI and obtained in controlled environments. Irregular’s tests, with 90 problems solved out of 288 in FrontierCyber, suggest less uniform performance than the announcements imply.²⁴ The problem of fixes also remains open. According to t3n, more flaws found means more patches to write and test, and developers often cannot keep up, which would leave attackers more time. OpenAI itself speaks of a limited window to use these models before similar capabilities become available to attackers too.⁵³

This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.

Sources

Footnotes

1. GPT-6 Cyber, OpenAI automatizza la caccia ai bug: cosa cambia per le imprese, AI4Business, 2026-11-06 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

2. GPT-6 Cyber trova 81 bug in Firefox: capacità, limiti e rischi OT, Cyber Security 360, 2026-11-06 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³

3. OpenAI’s new GPT-6 Cyber AI finds 81 Firefox flaws, 948 Linux bugs, BleepingComputer, 2026-11-06 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. OpenAI Launches GPT-6 Cyber After AI Helps Find 81 Firefox Flaws, 948 Linux Exploits, The Hacker News, 2026-11-06 ↩ ↩² ↩³ ↩⁴ ↩⁵

5. Dank KI: OpenAI findet 948 Lücken im Linux-Kernel – und schließt sie auch gleich, t3n, 2026-11-06 ↩ ↩²