48.5% beats 32.1%. OpenAI’s hacker just got sharper.
OpenAI released GPT-6 Cyber on November 5, 2026 — a GPT-6 Astra variant built to find software flaws and write working exploits. On ExploitGym, the benchmark that tests whether a model can exploit a specified vulnerability, it scored 48.5%, up from 32.1% for GPT-5.6 Cyber. In a controlled research run it built working privilege-escalation exploits for 948 Linux kernel vulnerabilities, 588 of them previously unreported, and found 81 previously unknown Firefox flaws that Mozilla rated high severity. OpenAI now rates the model Critical for cybersecurity under its Preparedness Framework, up from High — its top tier.
The storyline
Hook: 48.5% beats 32.1%. OpenAI’s hacker just got sharper.
Context: OpenAI released GPT-6 Cyber on November 5. It finds flaws and writes working exploits — and it is not available to the public. Access is restricted to approved Daybreak Red customers, through the API and Codex only.
Stakes: It built working exploits for 948 Linux flaws — nearly six times the old model’s 166. 588 were never reported before, and Mozilla rated 81 Firefox flaws high severity. It stays in bounds far better than its predecessor: forbidden-system touches in the honeypot test fell to 0.07%, from 5.51%. But there is a catch — it stops on warnings 73% of the time, under GPT-6 Astra’s 78%.
Payoff: Sharper hacker, tighter lock. You cannot try it — only vetted Daybreak Red customers get access.
Watch next
Kimi K3.1 Pro: the biggest open-weight AI release yet — Moonshot published the full weights on November 4.
How we verified this
Every figure above comes from OpenAI’s own November 5 announcement and the GPT-6 Cyber system card, cross-checked against independent coverage (The Neuron, AI Agent Update). All benchmark and exploit counts are OpenAI’s own reported figures — no outside lab has independently audited them yet. The 73% vs 78% stop-compliance comparison is against GPT-6 Astra, per the system card.
Links
- Instagram reel: https://www.instagram.com/p/DfcHbUpg9NC/
- OpenAI announcement: https://openai.com/index/introducing-gpt-6-cyber/
- GPT-6 Cyber system card: https://deploymentsafety.openai.com/gpt-6-cyber/gpt-6-cyber.pdf
- The Neuron — GPT-6 Cyber: OpenAI’s Hacking AI and Its Safety Limits: https://www.theneuron.ai/news/openais-gpt-6-cyber-can-hack-better-can-it-stay-within-the-rules/
- AI Agent Update — OpenAI ships GPT-6 Cyber, rated Critical: https://aiagentupdate.com/post/openai-gpt6-cyber-critical-2026-11-06
- https://www.instagram.com/simpleaiexplains/
- https://buymeacoffee.com/simpleaiexplains
Leave a Reply