OpenAI's GPT-6 Cyber can find and exploit flaws — this week showed why that matters

What OpenAI released

On November 5, 2026, OpenAI released GPT-6 Cyber, a model for cybersecurity research. Mayfield's newsletter confirms the release, describing it as a specialized model for advanced cybersecurity research. m:…0313 m:…9315

The capability numbers come from OpenAI itself, as reported by the Labellerr blog. Researchers used GPT-6 Cyber to find 81 new, high-severity flaws in the Firefox browser. The model also helped build working exploits for 948 Linux kernel flaws, 588 of which were bugs that no one had found before. In another test, the model found and confirmed a critical remote code execution flaw in Brotli. m:…0313

On ExploitGym, GPT-6 Cyber scored 48.5 percent. The older GPT-5.6 Cyber scored 32.1 percent on the same test. m:…0313

The highest rating, and a locked door

GPT-6 Cyber sits in the Critical tier for cyber skills, which is the highest level on the scale. A Critical rating means the model can find and build working zero-day exploits for real critical systems, and can also carry out cyberattacks from start to finish. m:…0313

Access is limited. The model is first available to a small group of approved customers in Daybreak Red. OpenAI also cancelled a planned GPT-6.1 Astra release, since that model was better at chasing goals but worse at staying within bounds. m:…0313

Four flaws from the same week

The same week brought four disclosures of the kind of flaw this model is built to find. None of the four disclosures says AI found the flaw. They are worth reading next to the release because they show what is sitting in software people actually run. m:…2680 m:…8768 m:…4020 m:…4019

The most severe is a LatePoint flaw, rated CVSS 9.1 Critical. An unauthenticated attacker can reset the password of a LatePoint customer account by sending the literal string IS NOT NULL as a password-reset token. The plugin has 100,000 active installs. The 5.7.8 release removes the literal-string branch in the query builder. No exploitation in the wild is reported, and the flaw is not in CISA KEV as of November 7, 2026. m:…2680

A Cronicle flaw, rated CVSS 9.0 Critical, lets a non-admin user with event privileges add undeclared keys to an event's params object. The job runner uppercases those keys into the child process environment, so the user can override LD_PRELOAD, PATH, NODE_OPTIONS and similar variables. Cronicle before 0.9.124 is affected. As with LatePoint, no exploitation in the wild is reported, and the flaw is not in CISA KEV as of November 7, 2026. m:…8768

Two WordPress plugins had SQL injection flaws, both rated CVSS 7.5 High by Wordfence, acting as the CNA. In wpForo, through version 3.2.2, any anonymous visitor who could load a forum page could reach a board lookup query that placed their input directly into SQL. WPScan reports about 20,000 active installations, and Wordfence credits the discovery to a researcher credited as UKO from Korea University. In Amelia Pro 9.9.1 or earlier, an unauthenticated visitor could inject SQL into an events query, with no account or user interaction needed. The free directory edition reports more than 90,000 active installations, the fix shipped in version 9.9.2 on November 4, 2026, and Wordfence credits researcher leon figueira of elitor.ai. m:…4020 m:…4019

What to take from the week

The pattern is the point. A model that can find and build working exploits is now gated behind vetting, while the flaws themselves are disclosed in public, in plugins with tens of thousands of installs, with fixes that only help if applied. The capability is restricted. The exposure is not. m:…0313 m:…2680 m:…4019

For developers, the week's work is unglamorous: check whether LatePoint, Cronicle, wpForo or Amelia run anywhere you are responsible for, and apply the fixed versions. The 5.7.8 and 9.9.2 releases and the 0.9.124 cutoff are the numbers that matter this week, and none of them depends on what you think of GPT-6 Cyber. m:…2680 m:…4019 m:…8768