OpenAI launches GPT-6 Cyber, rated Critical for cyber risk, after it found Firefox and Linux kernel flaws
OpenAI announced GPT-6 Cyber, a security-research model, and raised its cybersecurity risk rating for it from High to Critical. The figures come from OpenAI's own evaluations and have not been independently verified. OpenAI says the model scored 48.5% on ExploitGym, against 32.1% for GPT-5.6 Cyber. It also says researchers used it to reproduce 1,724 Linux kernel vulnerabilities and build working exploits for 948 of them. Of those, 304 critical findings were disclosed to maintainers with candidate patches. In Firefox, it produced reproducible evidence for 81 new flaws that Mozilla rated high severity. Access is limited to approved Daybreak Red customers, and the article reports no exploitation in the wild.
Detalhes técnicos
- In the Firefox work, the model also showed that 30 known issues were worse than first reported. Mozilla began landing fixes after the reports (securityaffairs.com).
- OpenAI also used the model, with others, to develop and validate exploits for a critical flaw in Brotli, a compression library used by Node.js. It says the flaw could lead to remote code execution in downstream applications (securityaffairs.com).
- In OpenAI's red-team campaigns, agents chained vulnerabilities to escape microVMs and run code on the underlying hosts (securityaffairs.com).
- Dragos tested the model on a mock power plant. It separated OT assets from IT noise and found two zero-days in engineering software that manages protective relay data (securityaffairs.com).
- In a test using known flaws in GitLab, PostgreSQL and Jenkins, GPT-6 Cyber completed 95% of tasks. GPT-6 Astra refused all of them (securityaffairs.com).
- OpenAI reports that the model avoided exploiting out-of-scope services in 99.93% of alignment tests, against 94.49% for GPT-5.6 Cyber. With the deployment prompt, it stayed in scope in 100% of tests (securityaffairs.com).
- OpenAI says it sees more misuse attempts. It monitored attempts to build exploits against live PHP servers during the disclosure of critical PHP vulnerabilities (securityaffairs.com).
- The article names no CVEs.
Remediação
- OpenAI recommends sandboxing, action review and human oversight for high-risk work.
- Safeguards it lists include Private Safety Processing, which reviews records kept in customer storage for 30 days. API customers on Zero Data Retention generally must use it.
- Daybreak Red API access requires IP allowlisting by default, with credential monitoring.
- OpenAI has also announced a $1B Daybreak for Frontline Defenders grant program and a Daybreak Red for Critical Infrastructure group (securityaffairs.com).
Linha do tempo
-
OpenAI launches GPT-6 Cyber, rated Critical for cyber risk, after it found Firefox and Linux kernel flaws
ler a análise desta fonte ocultar
OpenAI announced GPT-6 Cyber, a security-research model, and raised its cybersecurity risk rating for it from High to Critical. The figures come from OpenAI's own evaluations and have not been independently verified. OpenAI says the model scored 48.5% on ExploitGym, against 32.1% for GPT-5.6 Cyber. It also says researchers used it to reproduce 1,724 Linux kernel vulnerabilities and build working exploits for 948 of them. Of those, 304 critical findings were disclosed to maintainers with candidate patches. In Firefox, it produced reproducible evidence for 81 new flaws that Mozilla rated high severity. Access is limited to approved Daybreak Red customers, and the article reports no exploitation in the wild.
Technical Details
- In the Firefox work, the model also showed that 30 known issues were worse than first reported. Mozilla began landing fixes after the reports (securityaffairs.com).
- OpenAI also used the model, with others, to develop and validate exploits for a critical flaw in Brotli, a compression library used by Node.js. It says the flaw could lead to remote code execution in downstream applications (securityaffairs.com).
- In OpenAI's red-team campaigns, agents chained vulnerabilities to escape microVMs and run code on the underlying hosts (securityaffairs.com).
- Dragos tested the model on a mock power plant. It separated OT assets from IT noise and found two zero-days in engineering software that manages protective relay data (securityaffairs.com).
- In a test using known flaws in GitLab, PostgreSQL and Jenkins, GPT-6 Cyber completed 95% of tasks. GPT-6 Astra refused all of them (securityaffairs.com).
- OpenAI reports that the model avoided exploiting out-of-scope services in 99.93% of alignment tests, against 94.49% for GPT-5.6 Cyber. With the deployment prompt, it stayed in scope in 100% of tests (securityaffairs.com).
- OpenAI says it sees more misuse attempts. It monitored attempts to build exploits against live PHP servers during the disclosure of critical PHP vulnerabilities (securityaffairs.com).
- The article names no CVEs.
Remediation
- OpenAI recommends sandboxing, action review and human oversight for high-risk work.
- Safeguards it lists include Private Safety Processing, which reviews records kept in customer storage for 30 days. API customers on Zero Data Retention generally must use it.
- Daybreak Red API access requires IP allowlisting by default, with credential monitoring.
- OpenAI has also announced a $1B Daybreak for Frontline Defenders grant program and a Daybreak Red for Critical Infrastructure group (securityaffairs.com).