OpenAI launches GPT-6 Cyber for cybersecurity, stronger on exploits but pricier than rivals

OpenAI on November 5 released GPT-6 Cyber, a new model built for cybersecurity work that the company says finds software vulnerabilities far better than its predecessor while more reliably staying within the limits researchers set. The model costs more than rival offerings, however, and is initially open only to approved security teams. (Photo: file photo of the OpenAI logo.)

Big jump in exploit benchmarks

According to OpenAI, GPT-6 Cyber is a variant of GPT-6 Astra with additional training for security investigations. On ExploitGym, which tests whether a model can exploit a specified weakness, it scored 48.5%, against 32.1% for GPT-5.6 Cyber; on a web-application test conducted without access to source code, it found 40.3% of vulnerabilities, versus 32% for Astra.

In real research, researchers using GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948 of them; under the same setup, the previous generation produced exploits for 166. OpenAI said it disclosed 304 critical findings to Linux maintainers with proposed patches. The model also found 81 previously unknown high-severity vulnerabilities in the Firefox browser, plus evidence that 30 known issues deserved higher severity ratings; Mozilla has begun incorporating fixes.

Staying in scope: no attacks outside the assignment

OpenAI also stressed authorization boundaries. In its ExploitGym Honeypot test, tempting vulnerable services sit beside the assigned target, and the model must recognize that attacking them is out of scope. Without deployment safety prompts, GPT-6 Cyber avoided unauthorized attacks in 99.93% of tests, against 94.49% for GPT-5.6 Cyber; with the prompts, it stayed in scope in every tested case. The company rates its cybersecurity capability Critical under its Preparedness Framework, up from High for the previous generation.

Highest price, gated access

OpenAI's developer documentation lists GPT-6 Cyber at $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million. Input costs 60% more than GPT-5.6 Cyber, about twice Anthropic's Claude Mythos 5.1 and five times Google's Gemini 4 Argon Cyber Permissive. The model initially reaches selected customers through Daybreak Red, OpenAI's gated program for advanced authorized security work, which had enrolled more than 2,000 entities as of last month. The company also introduced Daybreak Red for Critical Infrastructure and highlighted its $1 billion program subsidizing access for resource-constrained defenders.

CodeRabbit CEO Harjot Gill said GPT-6 Cyber is "the strongest model we've tested for security review." VentureBeat noted, however, that the results come from OpenAI's own tests rather than a common independently administered comparison across vendors, and that deployment cost also depends on the compute and analyst review each investigation consumes.