In-scope scores are the lab's own

The OpenAI news feed dates GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents at Thu, 05 Nov 2026 09:00:00 GMT, which is 17:00 in Beijing on Nov. 5, inside this morning window. The fetched article text does not print a second clock, so this column uses that pubDate. The post introduces GPT-6 Cyber and says the cybersecurity class under the Preparedness Framework moves from High for GPT-5.6 Cyber to Critical. The class and the scores below are the company's own evaluations. A system card is linked. This column uses only the results stated in the announcement and does not turn the evaluations into a procedure.

On ExploitGym, the announcement gives GPT-6 Cyber 48.5 percent and GPT-5.6 Cyber 32.1 percent, about 16 percentage points higher and a 51 percent relative improvement. On a web evaluation the company built, it gives GPT-6 Cyber 40.3 percent recall and GPT-6 Astra 32.0 percent. On RemoteExploitBench, it says GPT-6 Astra with Daybreak Blue refused 100 percent of the tasks, while GPT-6 Cyber completed 95 percent, and the only failure was one that could not be made to work remotely. How the evaluation environment is built is not written as a procedure that can be followed, and this column does not add one.

The scope test is reported as a result. Without the deployment prompt, GPT-6 Cyber avoided out-of-scope services in 99.93 percent of tests. With the prompt, the figure is 100 percent. The corresponding figure for GPT-5.6 Cyber is given as 94.49 percent. The announcement calls this the result after training aimed at staying inside an authorized scope. How out-of-scope services were placed in the environment is not spelled out as a method.

The Firefox passage says that in one research campaign GPT-6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity, and that an additional 30 known issues were more severe than previously reported. The findings went to Mozilla with proposed patches, and Mozilla began landing fixes. The Linux kernel passage links the article already published and repeats 1,724 reproductions and 948 exploits. Those two counts are not a new cost or patch ledger in this post. The cost and the patch accounting stay with the earlier article and are not repeated here. Access for now is the API and Codex for a select group of approved Daybreak Red customers. The announcement says access will widen to all Daybreak Red customers later.

The announcement also names security teams at Cisco, Calif, and CodeRabbit as using it for vulnerability research and code review. Calif is copied as the page prints it. The page does not explain the name. The lab says it has also used the model on its own infrastructure and links another post. The details are not in this announcement. The Mozilla passage says fixes began landing after the reports arrived. Those are statements of use and of what was done with the reports, not a list of the flaws.

What this column records is the announcement's own scores, class, and access. The 48.5 percent, the 99.93 percent, and the 81 have no independent replication. The image is the announcement's social card. It shows the title, not a score table. Chain-of-thought excerpts and particular exploit methods are not relayed.