GPT-6 Cyber for vulnerability research

OpenAI says GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber.[1] It reproduced 1,724 vulnerabilities and developed working exploits for 948 of them.[1] OpenAI also reported a total run cost of $1,213,430 in API credits, or $1,279 per LPE exploit.[6] These figures describe the reported evaluation and research runs; they do not, on their own, establish what a team should expect for its own workload.

OpenAI says that without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests, compared with 94.49% for GPT-5.6 Cyber.[1] With deployment safety prompts, it stayed within bounds in all tests.[1] OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.[1] The company says it added safeguards designed to detect and disrupt misuse, including checks to block high-risk activity involving critical systems.[1]

OpenAI says GPT-6 Cyber produced evidence and reproducible tests for 81 novel Firefox vulnerabilities that Mozilla classified as high severity; Mozilla began landing fixes shortly after receiving the reports.[1] In an initial evaluation against GPT-6 Astra, Astra refused 100% of tasks, while Cyber completed 95%; Cyber’s only failure was an exploit it could not get to work remotely.[1] OpenAI also says Cyber achieved 40.3% vulnerability recall, compared with 32.0% for Astra.[1] These results do not show how much the models accelerate enterprise patch deployment or reduce exposure.[5]

OpenAI says GPT-6 Cyber is available to select Daybreak Red customers and plans to expand access to all Daybreak Red customers in the coming weeks.[2][3] OpenAI describes Daybreak Red as supporting work such as attack simulation and red-teaming.[4] Security teams at Mozilla, Palo Alto Networks, Calif and Cisco, as well as OpenAI’s own team, are already using the model, according to OpenAI.[2] The company also says it used GPT-6 Cyber to find and fix weaknesses in its infrastructure.[1] A reported concern is that stronger vulnerability-finding could leave more bugs exposed if they are found but not fixed.[5]