OpenAI 宣布推出名为 GPT-6 Cyber 的新网络安全模型,旨在推进“对齐的网络安全智能体”的前沿发展 134。该模型的核心主张在于其不仅具备更强的攻击面识别能力,同时在防止未经授权地攻击其他公司方面表现出更优的对齐特性 14。通过自动化流程,该模型在一次为期 30 小时的运行中,成功生成了针对 Linux 内核本地权限提升(Root Escalation)漏洞的利用代码,数量达到 948 个 1234。
除了针对操作系统的测试外,相关研究还显示该 AI 模型协助发现了 Firefox 浏览器中的 81 个安全缺陷 4。这一进展凸显了人工智能在自动化软件漏洞挖掘方面的巨大潜力,但也引发了业界对于双用途风险的广泛讨论 13。媒体指出,尽管 OpenAI 强调模型的防御性目标,但其生成近一千个 Linux 漏洞利用代码的能力仍引起了关于 AI 驱动网络攻击风险增加的担忧 3。目前报道未提供具体的发布日期或详细的技术基准数据,仅确认了上述关键测试结果和模型定位 1234。
OpenAI has announced the release of its new cybersecurity model, GPT-6 Cyber 134. The company states that this model represents the frontier of aligned cybersecurity agents and is better at not hacking other companies 14. Despite these design principles focusing on defensive tasks and alignment, reports indicate the model demonstrated significant dual-use capabilities during testing 13.
During a 30-hour run, GPT-6 Cyber generated exploit code for 948 Linux root privilege escalation vulnerabilities 1234. Hacker News cited a report by Phoronix stating that OpenAI found these local privilege escalation bugs in the Linux kernel 2. Axios noted that the number of flaws approached nearly 1,000, highlighting the model's ability to automate the discovery of software security weaknesses 3. In addition to the Linux findings, The Hacker News reported that the AI helped discover 81 flaws in Firefox 4. These developments have sparked discussions regarding the potential for AI-driven cyberattacks, even as the model aims to strengthen defense 13.
评论
还没有评论,欢迎留下第一条。