OpenAI announced GPT-6 Cyber on Nov. 5 and said the model found 81 previously unknown high-severity vulnerabilities in Firefox. The same model, the company reported, built working local privilege-escalation exploits for 948 distinct Linux kernel flaws. Neither capability is open to the public, because access is restricted to approved security customers.
AI, explained every day: The Deep View is the go-to daily newsletter for 600,000+ founders, knowledge workers and professionals who want to stay up to date with artificial intelligence.
We may earn a commission if you sign up.
OpenAI rated GPT-6 Cyber “Critical” for cybersecurity capability under its Preparedness Framework, a step up from the “High” rating given to GPT-5.6 Cyber. The company framed the release around a narrow window for defenders: in its telling, frontier models can still be used to secure software before attackers gain comparable tools, but not for long. The Firefox count is the part of the announcement an outsider can partly test, because Mozilla publishes its fixes in public advisories.
Mozilla’s patches for the 81 Firefox flaws
BleepingComputer’s Lawrence Abrams reported on Nov. 6 that GPT-6 Cyber helped develop patches as well as find the bugs, and that Mozilla has already released many of them under advisory MFSA 2026-118. Help Net Security’s Anamarija Pogorelec added in her own write-up that Mozilla classified all 81 flaws as high severity and that the model contributed proposed fixes.
Brian Grinstead, a Distinguished Engineer at Firefox, was the only Mozilla voice quoted in coverage. He said Firefox had already been hardened with several generations of frontier models, yet GPT-6 Cyber still turned up previously unknown high-severity vulnerabilities. According to Digital Trends, he also said the resulting fixes have reached millions of users, and the model showed that 30 already-known bugs were more dangerous than first reported. Anyone with automatic updates switched off should install the newest Firefox build by hand.
The Firefox 159 advisory record
Mozilla’s own advisory is the primary record behind the patches, and it is much broader than the 81. MFSA 2026-118, announced Oct. 27 for Firefox 159, rates its overall impact High and lists 167 entries by a tally of the page: 108 high, 52 moderate and 7 low. Mozilla says it no longer folds internally found memory-safety bugs into one CVE, so each bug now gets a separate entry.
The advisory credits Amy Burnett of OpenAI on 7 entries and OpenAI Codex Security, under the handle kylebot, on 34 more. It never mentions GPT-6 Cyber or any other OpenAI model, and it predates the announcement by nine days. Matching the 81 flaws to specific advisory entries is therefore not possible from the public page alone. The figure rests on OpenAI’s account, with Grinstead confirming the broad finding that the bugs were new and serious.
948 Linux kernel exploits and 304 disclosures to maintainers
On Linux, OpenAI’s research as relayed by BleepingComputer says the model flagged 2,308 potential flaws, reproduced 1,724 of them and developed exploits for 948 distinct vulnerabilities, the kind that let an unprivileged user gain root access. The Next Web adds that OpenAI has disclosed 304 critical findings to maintainers along with proposed patches, a pipeline that keeps the exploit work tied to fixes.
On the ExploitGym benchmark, GPT-6 Cyber exploited 48.5% of vulnerabilities against 32.1% for GPT-5.6 Cyber. Chris Hughes, vice president of security strategy at Noma Security, pointed to OpenAI’s Patch the Planet dashboard, which he said showed fewer than 25% of the flaws found so far as fixed, and cautioned that at the scale these models operate even a small fraction of misbehavior means actions against systems nobody authorized.
Dragos tested the model on a mock power plant and reported two zero-day flaws in engineering software that manages protective relay data. Rob T. Lee, chief AI officer at Dragos, said the skill floor is dropping for the reconnaissance and vulnerability research that precede an attack.
Daybreak Red access and the safety claims
Access runs through Daybreak Red, OpenAI’s tier for exploit development and authorized security testing, via the API and Codex, with plans to open the model to every Daybreak Red customer soon. Fouad Matin, who leads OpenAI’s cyber preparedness and safety team, said the model follows instructions more than 99 percent of the time. In an off-limits-services test it avoided those services in 99.93% of runs without its deployment safety prompt, against 94.49% for GPT-5.6 Cyber, and stayed in bounds every time with the prompt.
BleepingComputer placed the launch among a run of AI-assisted vulnerability finds. In October, researchers from OpenAI, Calif and Wiz used AI models to find two critical PHP flaws dubbed PHP2Shell, and a security researcher ran roughly 17,600 authentication tests with AI agents to uncover an OpenSSH authentication bypass called sshole. Each result moved from lab curiosity to patchable bug report, which is the same path the Firefox findings are said to have followed.
The open question is verification. OpenAI has promised to extend its Daybreak programs to organizations of all sizes, independent researchers and open-source maintainers, but until the 81 flaws are mapped one by one to published Mozilla advisory entries, the count remains OpenAI’s own, backed by Grinstead’s account that the bugs were new and serious.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
AI, explained every day: Get The Deep View →
We may earn a commission if you sign up.
More from Morning Overview
- Shaq spent $1.3 million on three Rolls-Royce Phantoms to prove a point to a salesman
- Over 20 trillion gallons of rain fell on the Four Corners states and Nevada in two weeks
- A handful of car transmissions are so tough that mechanics say they almost never die
- Two Kentucky residents died when a small plane crashed at Missouri’s Lake of the Ozarks, and the NTSB is investigating
