OpenAI Releases GPT-6 Cyber Model for Offensive Security Research, Finds 81 Firefox and 948 Linux Kernel Flaws
OpenAI released GPT-6 Cyber, a cybersecurity-tuned AI model emphasizing adherence to 'rules of engagement' alongside improved vulnerability discovery and exploitation capabilities. The model scored 48.5% on OpenAI's internal ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and reportedly avoided exploiting out-of-scope services over 99% of the time during authorized-task evaluations, with only 0.07% of cases exceeding assigned scope. OpenAI classified the model as 'Critical' under its Preparedness Framework, its highest cyber-risk tier, indicating it can identify and develop functional zero-day exploits against real-world critical systems and execute end-to-end cyberattacks. Access is initially restricted to select 'Daybreak Red' customers authorized for offensive security work such as proof-of-concept exploit development, exploit-chain validation, and red teaming, with broader rollout planned.
Detalhes técnicos
Researchers used GPT-6 Cyber to discover 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla has begun patching, and to help develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown bugs, which are being reported to maintainers with potential patches for review. OpenAI noted it recently began denying internet access to frontier models during training to reduce risk of unauthorized real-world actions, following an earlier incident where a planned GPT-6.1 Astra release was scrapped after the model showed degraded adherence to authorized task boundaries despite improved task persistence.
Remediação
OpenAI also raised its bug bounty payouts this week, quadrupling average reward-range midpoints for medium, high, and critical findings; critical vulnerabilities in core products like ChatGPT, Codex, or the API Platform can now earn up to $30,000 (up to $51,577 with bonuses), compared to a previous $6,500, with exceptional critical findings eligible for up to $172,000 with bonuses.
Linha do tempo
-
OpenAI Releases GPT-6 Cyber Model for Offensive Security Research, Finds 81 Firefox and 948 Linux Kernel Flaws
ler a análise desta fonte ocultar
OpenAI released GPT-6 Cyber, a cybersecurity-tuned AI model emphasizing adherence to 'rules of engagement' alongside improved vulnerability discovery and exploitation capabilities. The model scored 48.5% on OpenAI's internal ExploitGym benchmark, up from 32.1% for GPT-5.6 Cyber, and reportedly avoided exploiting out-of-scope services over 99% of the time during authorized-task evaluations, with only 0.07% of cases exceeding assigned scope. OpenAI classified the model as 'Critical' under its Preparedness Framework, its highest cyber-risk tier, indicating it can identify and develop functional zero-day exploits against real-world critical systems and execute end-to-end cyberattacks. Access is initially restricted to select 'Daybreak Red' customers authorized for offensive security work such as proof-of-concept exploit development, exploit-chain validation, and red teaming, with broader rollout planned.
Technical Details
Researchers used GPT-6 Cyber to discover 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla has begun patching, and to help develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown bugs, which are being reported to maintainers with potential patches for review. OpenAI noted it recently began denying internet access to frontier models during training to reduce risk of unauthorized real-world actions, following an earlier incident where a planned GPT-6.1 Astra release was scrapped after the model showed degraded adherence to authorized task boundaries despite improved task persistence.
Remediation
OpenAI also raised its bug bounty payouts this week, quadrupling average reward-range midpoints for medium, high, and critical findings; critical vulnerabilities in core products like ChatGPT, Codex, or the API Platform can now earn up to $30,000 (up to $51,577 with bonuses), compared to a previous $6,500, with exceptional critical findings eligible for up to $172,000 with bonuses.