人工智能时代的网络安全像现代战争一样高度不对称，少数攻击者借助最新模型就能突破强大防线，而防守方必须每次都正确。Anthropic的Mythos Preview和OpenAI的5.5-Cyber已被限制向可信企业开放，但更广泛的风险正在上升，因为“AI必须对抗AI”。

已有数据表明攻击速度和频率在急剧恶化：CrowdStrike称，2025年的AI增强攻击比前一年增长了89%；Palo Alto Networks发现，配备AI后，最快四分之一攻击者在2025年只需略多于1小时就能从入侵系统中窃取数据，而2024年接近5小时。与此同时，漏洞报告数持续飙升，业内甚至预计年报漏洞数可能从当前水平跃升10倍至480,000。

AI既在武装攻击者，也在帮助防御者。Mozilla称，早期版Mythos帮助其识别出Firefox新版本中的271个漏洞，且能够发现人类可识别的全部漏洞；Cisco则在5月12日宣布推出开源分步指南，以帮助企业构建安全“harnesses”来增强防御。尽管如此，业内对执行力的评分仍只有B+，因为新威胁仍在快速演化，且“最佳答案”尚未出现。

Cyber-security in the age of artificial intelligence is highly asymmetric, like modern warfare: a small number of attackers armed with the latest models can breach powerful defences, while defenders must be right every time. Anthropic’s Mythos Preview and OpenAI’s 5.5-Cyber have been limited to trusted firms, but the wider risk keeps rising because “AI has to fight AI.”

The data already show a sharp deterioration in attack speed and frequency: CrowdStrike says AI-enhanced attacks rose by 89% in 2025 from the year before; Palo Alto Networks found that, with AI, the fastest quartile of attackers could steal data from a system they had broken into in just over 1 hour in 2025, down from almost 5 hours in 2024. At the same time, vulnerability reports are surging, and the industry even expects annual CVEs could rise ten-fold to 480,000.

AI is both arming attackers and helping defenders. Mozilla said an early version of Mythos helped it identify 271 vulnerabilities in a new Firefox release and could find every bug a human could; Cisco announced on May 12th an open-source, step-by-step guide for building cyber-security harnesses to help companies use any model to strengthen defences. Even so, the industry’s execution is judged only a B+, because the threats are still evolving quickly and no “magic bullet” has emerged.

Source: The war between businesses and hackers enters a perilous new phase

Subtitle: AI agents present novel dangers

Dateline: 5月 14, 2026 11:24 上午 | LOS ANGELES