• OpenAI has launched GPT-6 Cyber, elevating its cybersecurity capability classification to Critical from High for GPT-5.6 Cyber.
  • The model reportedly identified 81 previously unknown high-severity Firefox vulnerabilities and helped develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown issues.
  • GPT-6 Cyber is rolling out to select Daybreak Red customers, with broader access planned over the coming weeks; OpenAI warns of increasing attempts by malicious actors to misuse cybersecurity models.

A New Threshold in Cyber Capability

OpenAI has introduced GPT-6 Cyber, upgrading its cybersecurity capability classification to Critical from High for GPT-5.6 Cyber. The model reportedly identified 81 previously unknown high-severity Firefox vulnerabilities and helped identify and develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown issues. GPT-6 Cyber is rolling out to select Daybreak Red customers, with broader access planned over the coming weeks, according to people familiar with the matter.

The disclosure comes amid a broader push by OpenAI to position its cybersecurity offerings at the forefront of the market. The company also warned of increasing attempts by malicious actors to misuse cybersecurity models. A spokesperson for OpenAI declined to comment beyond the public announcement.

Verified Results, Unverified Claims

While the Firefox figure and the Critical label for this specific model remain unverified, the surrounding facts have been corroborated. OpenAI published a Linux kernel study on November 4, 2026. The company's pipeline analyzed 2,308 candidate kernel findings and reproduced 1,724 vulnerabilities. It produced 948 working local privilege escalation exploits, which give an unprivileged user root access. The run used an "upcoming cyber model" and cost $1,213,430 in API credits, about $1,279 per exploit.

The 948 figure counts distinct exploits and patches after variant analysis. It does not mean 948 flaws in a single component. The post does not give a "588 previously unknown" figure. It does say that while the scan ran, 61 of the exploitable bugs had been patched upstream, 43 had staging patches, and 299 were being discussed publicly. A July run with GPT-5.6 Sol found 166 such bugs. The October run found more than five times as many despite three months of kernel hardening. The agents used at least 10 novel exploitation techniques, which outside kernel experts also hadn't seen, according to a person familiar with the matter.

Racing to Keep Pace

The cyber push is also a commercial strategy. Anthropic is reportedly winning enterprise customers, and OpenAI is racing to catch up. Tech stocks sold off sharply in late October, and Meta (NASDAQ:META) received a credit warning, so investors are scrutinizing AI spending. OpenAI's own math shows how cheap offense has become. Its pipeline's output would have been worth a hypothetical $18.9 million to $66.3 million in historical bug bounties, 16 to 55 times its cost. That favors attackers and defenders alike, and OpenAI says unpatched systems with a low-privilege account should be treated as equivalent to root access for a well-resourced attacker.

OpenAI is an AI lab whose products include ChatGPT, Codex, the API and the Daybreak cyber program. Since October 2025 it has run as OpenAI Group PBC, a public benefit corporation controlled by the nonprofit OpenAI Foundation. The Foundation holds about 26% and Sam Altman sits on its board. Its SEC filing showed first-half 2026 revenue of $16.3 billion. Bloomberg-sourced reports put annualized revenue near $50 billion at the end of September. A market-news outlet cautioned that run-rate figures are unverified until accounts are published. The WSJ, via Futurism, said Q2 sales grew only 18% to $6.7 billion, while Anthropic doubled revenue. That's a contested comparison.

OpenAI cut its latest fundraising target to $20 billion from $30 billion, at a $1.25 trillion valuation instead of $1.4 trillion. About $11 billion was committed, and the round had not closed. In March it had announced $122 billion in committed capital at an $852 billion valuation. Its IPO has been pushed to 2027 at the earliest. Uber (NYSE:UBER)'s Jill Hazelbaker is reportedly joining as Chief Corporate Affairs Officer. The Foundation hired a new CFO and several program heads in 2026.

A Double-Edged Sword

Defenders and maintainers gain faster discovery and patching, but the volume of reports strains them. OpenAI says its triage bottleneck was prioritization, not discovery. Linux underpins Android, cloud and supercomputers, so patch discipline matters more. The headline's warning about malicious actors fits OpenAI's stated rationale for restricting access. There was no independent public reaction to this specific launch.

The EU's cybersecurity agency ENISA received GPT-5.5-Cyber in May and later GPT-6 Astra. It is also testing Anthropic's Mythos 5. An EU spokesman said "a new wave of powerful models are coming" and that Brussels is coordinating with the US. The EU's high-risk AI obligations were delayed to December 2027, though transparency rules took effect in August 2026. OpenAI extended cyber access to Ukraine for civilian defense in September. Access is gated by verification rather than law, so company policy is currently the main control.

OpenAI's history with cyber models includes the February 2026 introduction of Trusted Access for Cyber, the June 2026 launch of GPT-5.5-Cyber, and the August 2026 launch of GPT-5.6-Cyber with Daybreak Red. On September 3, 2026, GPT-6 Astra became OpenAI's first model to reach its internal Critical cyber threshold. In November 2026, GPT-6.1 Terra was reported to meet Critical too. Precedent includes Anthropic's Mythos and Project Glasswing, with more than 10,000 high- or critical-severity bugs found in a month by about 50 partners. Claude also found 22 Firefox vulnerabilities in early 2026.

Expect broader GPT-6 Cyber access over weeks, more upstream Linux patches, and more rescans of other critical software. OpenAI recommends rescanning with each new model generation. Long term, OpenAI predicts frontier models will find essentially any vulnerability in a project within a few years. It says that raises the importance of secure-by-default code, smaller attack surfaces and formal verification. That is a company view, not independent analysis. Scale and automation can produce wrong results. In an early run, 20 of 200 "successes" used a different 0-day than assigned.

Anthropic reported an unreleased model finding a Chrome remote code execution bug and a Safari exploit chain. Anthropic is preparing an IPO, and its S-1 reportedly showed positive cash flow in the first half of 2026.

Correction: An earlier version of this article misstated the number of previously unknown Linux kernel vulnerabilities. OpenAI's study did not specify a figure of 588 previously unknown issues; that number remains unconfirmed.