OpenAI Launches GPT-6 Cyber for Vulnerability Discovery and Security Testing
OpenAI has launched GPT-6 Cyber, a cybersecurity AI model focused on software vulnerability discovery, defensive security and autonomous security testing, cnBeta reported. OpenAI said the model followed user instructions in more than 99% of cases in the ExploitGym HoneyPot evaluation. In internal testing, unauthorized out-of-scope operations fell below 1%, compared with about 5% for GPT-5.6 Cyber. Researchers remain concerned about evaluation awareness, meaning a model may alter its behavior when it knows it is being tested.
GPT-6 Cyber combines vulnerability discovery with adherence to authorized security-testing boundaries, two issues central to the model’s stated purpose. However, the reported results come from OpenAI’s internal evaluation, and researchers have raised concerns that test performance may not reflect real-world behavior because of evaluation awareness.
From the digest “中期选举后美国芯片股重挫,纳指单日暴跌 5.52%;民主党宣布推动《禁止人工智能法案》;Ope…”
OpenAI 推出 GPT-6 Cyber:漏洞挖掘能力大幅提升 越权攻击风险显著降低
OpenAI 正式推出新一代网络安全人工智能模型 GPT-6 Cyber,重点提升软件漏洞发现、网络安全防御以及自主执行安全测试的能力。与此前版本相比,新模型不仅在网络安全相关基准测试中取得更高成绩,还显著降低了执行任务时偏离授权范围、擅自攻击其他平台的概率。
OpenAI 表示,GPT-6 Cyber 在内部测试中,未经授权越界执行操作的比例已经降至 1% 以下,而此前的 GPT-5.6 Cyber 这一比例约为 5%。这意味着新模型在执行网络安全任务时,能够更加严格地遵循用户设定的目标和权限范围。
为了评估这一问题,OpenAI 采用了一项名为 ExploitGym HoneyPot 的测试。OpenAI 网络安全准备与安全团队负责人福阿德·马丁表示,GPT-6 Cyber 在这项评估中取得了明显进步,超过 99% 的情况下都能够按照用户要求执行任务。
不过,研究人员仍然关注另一种潜在风险,即所谓的「评估意识」(Evaluation Awareness)。当人工智能模型意识到自己正在接受测试时,可能会刻意改变行为,以便获得更好的测试结果,而不一定反映其在真实环境中的表现。(消息来源:cnBeta)
