GPT-6 Cyber: OpenAI's New AI Found 81 Hidden Firefox Security Flaws

Try Unlimited Seedance 2.5 on OpenArt: https://openart.ai/home?utm_source=youtube&utm_medium=influencer&utm_campaign=infl-youtube--na-acq-web&ref=BitBiasedAI-sd25 OpenAI's GPT-6 Cyber reportedly generated **948 working Linux kernel exploits in just ten days**, discovered 81 previously unknown high-severity Firefox vulnerabilities, and scored 48.5% on ExploitGym. The estimated cost? Around **$1,280 per working exploit**. Those numbers suggest a major shift in AI-powered cybersecurity. But there's a catch: the headline benchmarks, the actual capabilities, and what independent researchers can verify tell three different stories. OpenAI introduced GPT-6 Cyber on November 5 as a specialized cybersecurity variant of GPT-6 Astra, designed to discover vulnerabilities, reproduce security flaws, develop working exploits, and help researchers test defenses. Unlike ordinary ChatGPT models, GPT-6 Cyber is restricted to vetted customers through OpenAI's Daybreak Red program, with identity verification, hardware security keys, and additional access controls. The Linux experiment is particularly revealing. According to OpenAI, a multi-agent pipeline scanned approximately 5,000 security-relevant locations in the Linux kernel, identified 2,308 candidate crashes, confirmed 1,724 bugs, and produced 948 working exploits. OpenAI reports that 588 of those exploits targeted previously unknown vulnerabilities. The entire experiment reportedly consumed $1,213,430 in API credits. Compared with traditional vulnerability research and exploit-development costs, that raises an uncomfortable question: **What happens when discovering serious software vulnerabilities becomes dramatically cheaper than fixing them?** Then there's Firefox. OpenAI says GPT-6 Cyber identified 81 previously unknown high-severity Firefox vulnerabilities and supplied reproducible evidence to Mozilla. Firefox 159's security advisory provides evidence of substantial security fixes, but the public record doesn't independently connect all 81 reported discoveries to GPT-6 Cyber. The comparison with Anthropic is also important. Mozilla previously reported protections for 271 vulnerabilities discovered with Anthropic's Mythos Preview. Different testing conditions and counting methods make direct comparisons difficult, but the broader trend is unmistakable: AI-assisted vulnerability discovery is creating an enormous new workload for software maintainers. And the benchmark story gets even more interesting. However, the general-purpose GPT-6 Astra reportedly achieved 42.4% without the same cyber specialization. That makes the additional improvement from cyber-specific training considerably smaller than the headline comparison suggests. Google's Gemini 4 Carbon is another serious competitor, with an independently unverified leaderboard listing it at 53.5% on ExploitGym. Meanwhile, OpenAI's general-purpose model with safeguards enabled scored zero on certain exploit tasks because it refused to perform them. That's an important distinction between lacking technical capability and being restricted from using it. OpenAI has also classified GPT-6 Cyber as **Critical** under its cybersecurity Preparedness Framework. Its reported honeypot evaluations show a dramatic reduction in unauthorized, out-of-scope actions compared with GPT-5.6 Cyber. But even a small failure rate becomes significant when automated systems operate at scale. We also examine OpenAI's reported Dragos power-grid security experiment, the $1 billion Daybreak initiative, and the growing competition between OpenAI, Google, and Anthropic to develop powerful cybersecurity AI while restricting potentially dangerous capabilities. The bigger question isn't simply whether GPT-6 Cyber is better at hacking. It's whether defenders can patch vulnerabilities quickly enough when AI systems can discover them at unprecedented scale, and what happens when similar capabilities become available outside tightly controlled programs. This breakdown separates OpenAI's reported results from independently verifiable evidence, examines the limitations behind the benchmarks, and explores what GPT-6 Cyber could mean for cybersecurity researchers, open-source maintainers, businesses, and the future of AI-powered security. Subscribe to BitBiased for evidence-driven coverage of OpenAI, artificial intelligence, cybersecurity, emerging technology, and the developments behind the headlines. CHAPTERS 00:00 GPT-6 Cyber: 948 Linux Exploits and the Real Story 01:18 What GPT-6 Cyber Is, and Who Isn't Allowed to Use It 03:26 Seedance 2.5 by OpenArt 05:27 948 Linux Exploits in Ten Days, for About $1,280 Each 08:35 Firefox: 81 Bugs, and What the Public Record Can Confirm 11:09 What a 48.5% on ExploitGym Actually Means 14:06 "Critical" and the Safety Numbers 15:43 Who Benefits First, Defenders or Attackers? 17:32 What to Take From GPT-6 Cyber #OpenAI #GPT6Cyber #Cybersecurity #AI #LinuxSecurity