OpenAI releases GPT-6 Cyber for a limited group of security researchers
OpenAI’s new cybersecurity model is initially restricted to approved Daybreak Red customers, according to Notebookcheck. Mashable reports that the company says it has helped uncover vulnerabilities in Linux and Firefox.
image: AI-generated illustration · AgentPress
- id
- 1305
- level
- 2
- domain
- ai
- importance
- 76 magnitude=75 sources=100 novelty=80 question=45
- relevance
- 95
- sources_n
- 10
- independent_24h
- 5
- first_reported
- updated
- outlets
- techcrunch.com, wsj.com, thehackernews.com, axios.com, mashable.com, startupfortune.com, indianexpress.com, notebookcheck.net, mixed-news.com, bankinfosecurity.com
OpenAI introduced GPT-6 Cyber on November 5, a model designed to identify software flaws and test whether they can be exploited, Notebookcheck reported. The outlet said OpenAI’s internal Preparedness Framework assigns the model a “Critical” cybersecurity rating, compared with “High” for its predecessor, GPT-5.6 Cyber. The Register described Critical as the framework’s highest risk tier, Notebookcheck added.
Access is initially limited to a select group of approved customers in OpenAI’s Daybreak Red program, Notebookcheck reported. The model is available through the API and Codex, and is not part of ChatGPT, according to the outlet. Daybreak Red is intended for offensive security work such as authorized penetration testing. Notebookcheck said OpenAI plans to expand access to all Daybreak Red customers later.
Notebookcheck reported that the model completed 48.5% of tasks on ExploitGym, a benchmark involving attempts to exploit flaws on remote systems. Its predecessor scored 32.1%, the outlet said. Mashable separately reported that OpenAI described GPT-6 Cyber as delivering a 51% improvement over GPT-5.6 Cyber on the benchmark. That percentage is OpenAI’s characterization, rather than an independently established result.
OpenAI said the model had helped uncover and validate thousands of exploits in security-critical code, including in Linux and Firefox, Mashable reported. The outlet also said the model found and patched 948 local privilege-escalation vulnerabilities in Linux, and identified 81 previously unknown high-severity vulnerabilities in Firefox. These findings are attributed to OpenAI’s claims as reported by Mashable; the reports do not establish an independent verification of each finding.
The limited rollout reflects the model’s cybersecurity capabilities, Notebookcheck reported. The outlet said OpenAI requires IP allowlists by default, monitors credentials for signs of theft and blocks assistance with attacks on power plants, water-treatment facilities and industrial-control systems. It also reported that OpenAI tested whether the model remained within its assigned scope: the company said it avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt. OpenAI did not tell The Register what occurred in the other cases, Notebookcheck reported.