The Decoder· Matthias Bastian·· 5 h ago
OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them
OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them
AI summary
OpenAI 推出专为挖掘软件安全漏洞训练的 GPT-6 Cyber，能发现缺陷、构建可用利用代码并建议补丁。OpenAI 称其在无额外安全指令时 99.93% 的测试未越界，前代为 94.49%；ExploitGym 得分 48.5%，前代 GPT-5.6 Cyber 为 32.1%。
Selection record
Threshold 76Media and individualsFirst 86Second 82
AdmittedSum of both 168 ≥ twice the threshold 152
- Source tier
- Media and individuals; this tier's threshold is 76
- Pre-filter
- passed：OpenAI发布GPT-6 Cyber安全模型及评测
- Why it was chosen
- 原文给出 GPT-6 Cyber 在 ExploitGym 与真实漏洞挖掘上的具体数字和算力成本，可据此判断安全专用模型的能力边界。
- Same event
A model scores each item twice, independently, against one written standard, out of 100. An item is admitted only when the two scores add up to twice the threshold. The threshold is set per source tier.
Source: The Decoder · the-decoder.com