OpenAI's GPT-6 Cyber claims 948 Linux exploits, 81 Firefox bugs

OpenAI has announced GPT-6 Cyber, a security-focused model it says built functional privilege-escalation exploits for 948 Linux kernel vulnerabilities and documented 81 high-severity Firefox flaws. The figures are worth an operator's attention — and worth reading carefully, because every one of them is OpenAI's own, and no CVE IDs are attached to any of them. This post is based on OpenAI's announcement as reported by SecurityWeek and CyberScoop; OpenAI's own report is the only source for the benchmark and vulnerability counts.

What OpenAI claims

• On ExploitGym, which tests whether a model can exploit a flaw in a remote target, GPT-6 Cyber scored 48.5%, up from GPT-5.6 Cyber's 32.1%. OpenAI rates it Critical under its Preparedness Framework, a step above the prior model's High.
• On the Linux kernel, the model "reproduced" 1,724 vulnerabilities and built working privilege-escalation exploits for 948 of them — against 166 for an earlier model using the same tooling. OpenAI says it reported 304 critical issues to maintainers with proposed fixes.
• On Firefox, it documented 81 new flaws that Mozilla rated high severity, each with evidence and a reproducible test. Mozilla has begun shipping patches.

Read the numbers skeptically

The AI-security beat is the most hype-prone one there is, and this is a vendor announcement. Four things the headline counts don't tell you:

1. No CVE IDs. OpenAI names no specific vulnerabilities and does not say whether any were assigned CVEs. A number you can't look up is a claim, not a finding.
2. "Reproduced" is not "discovered." The 948 figure sits on top of 1,724 reproduced kernel bugs — existing issues the model re-derived exploits for, not 948 fresh zero-days.
3. No independent verification of the Linux numbers is reported. Mozilla's role is confined to rating and patching the Firefox items.
4. The alignment scores are self-graded. The guardrail metrics below come from OpenAI's own tests.

The guardrails, as described

OpenAI says that in an alignment test seeded with off-limits decoy targets, the model avoided them 99.93% of the time without safety prompts, versus 94.49% for the prior version, and never touched them with the prompts in place. Industrial-control specialist Dragos tested the models in a controlled lab to refine safeguards against safety-critical targeting. Dragos framed the broader shift bluntly: "The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack." Access is gated to approved Daybreak Red customers through the API and Codex.

What to do

1. You can't run it — defenders don't get open access — so treat this as a forecast, not a tool. Plan for attackers who can automate kernel-exploit development at scale.
2. Prioritize Linux kernel privilege-escalation patching and Firefox updates now; if OpenAI's maintainer reports land, fixes will follow on those paths.
3. Hold the counts as vendor claims until CVEs are assigned and third parties confirm them. Forward the capability signal to your colleagues; don't forward the numbers as fact.

Context

This is OpenAI's fourth cyber model of 2026, each announced with a higher benchmark than the last. The capability trend is real and worth tracking. The disclosure practice — big counts, no lookup-able identifiers — is exactly what this newsroom hedges on, every time.