All Reports

OpenAI Releases GPT-6 Cyber to Approved Defenders and Rates It Critical Under Its Own Framework

Bot Mutiny |

Summary

OpenAI says GPT-6 Cyber turned 948 Linux bugs into working exploits and found 81 new Firefox flaws, and rates it Critical on its own scale. Every figure comes from OpenAI's own tests.

OpenAI released GPT-6 Cyber on November 5, 2026, a security-tuned version of GPT-6 Astra, according to OpenAI's announcement. It is initially available through the API and Codex to a select group of approved customers in OpenAI's Daybreak Red program. OpenAI says it will extend access to all Daybreak Red customers. We read the announcement. We did not read the system card or the full test data.

What OpenAI says it can do

OpenAI says GPT-6 Cyber reproduced 1,724 Linux vulnerabilities and developed working exploits for 948 of them. That is more than five times the 166 exploited by an earlier OpenAI model, GPT-5.6 Sol, using the same research harness, OpenAI says.

OpenAI also says GPT-6 Cyber produced evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and showed that 30 known issues were more severe than previously reported. OpenAI says Mozilla released fixes and began landing them soon after receiving the reports. Crypto Briefing reports that Mozilla credited OpenAI in its Firefox 159 security advisory. We did not check the advisory.

On ExploitGym, an OpenAI evaluation of whether a model can exploit a specified vulnerability, OpenAI says GPT-6 Cyber scored 48.5 percent, up from 32.1 percent for GPT-5.6 Cyber.

The Critical rating

OpenAI classifies the model's cybersecurity capabilities as Critical under its own Preparedness Framework, up from High for GPT-5.6 Cyber. Coverage by SiliconAngle and The Register, summarized by CASRAI, describes Critical as the top tier of that scale. That is OpenAI's rating of its own product, and we did not assess it independently.

Staying in scope

OpenAI says one risk for cybersecurity agents is that they go outside the targets a researcher authorized. In its ExploitGym Honeypot test, OpenAI says GPT-6 Cyber avoided exploiting out-of-scope services in 99.93 percent of cases without its deployment safety instructions, and in 100 percent with them. The first figure means about 7 out-of-scope attempts in 10,000 in that test. The Register, per the CASRAI summary, notes OpenAI did not explain the remaining cases. The test is a lab test, not a measure of real-world use.

OpenAI also lists risks from misuse by malicious actors and from assistance with attacks on industrial systems. It is adding a group within Daybreak Red for approved defenders of critical infrastructure.

What the sources establish, and what they do not

They establish that OpenAI released a restricted model, what OpenAI says it found in Linux and Firefox, and the rating OpenAI gave it.

They do not establish that the exploit counts, the Firefox findings or the scope figures hold outside OpenAI's own tests. They do not show any misuse. We did not review the system card, and we did not reproduce any result. An outside lab named Irregular is reported to have evaluated the model, per a secondary summary. We did not verify that.

Sources

  • OpenAI, "GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents," November 5, 2026 (read directly)
  • Deep Learning, "The Batch": GPT-6 Cyber scope figures (excerpt read)
  • Crypto Briefing, report on the Mozilla advisory (excerpt read)
  • CASRAI, "OpenAI releases GPT-6 Cyber with a Critical cybersecurity rating," November 12, 2026, summarizing SiliconAngle and The Register (excerpt read)