Google's Threat Intelligence Group (GTIG) has reported the disruption of a large-scale cyber operation that utilized large language models to identify and exploit a previously unknown software vulnerability.

Google 的威脅情報小組 (GTIG) 報告指出，已瓦解一項利用大型語言模型來識別並利用先前未知軟體漏洞的大規模網路行動。

The operation in question targeted a web-based system administration tool, leveraging a semantic logic flaw to circumvent two-factor authentication. GTIG identified the use of artificial intelligence through the presence of 'hallucinated' CVSS scores and textbook formatting within the Python scripts, which are characteristic of LLM training data. While the specific model employed remains unidentified, Google has indicated that its own Gemini model was likely not utilized. This incident aligns with broader observations that criminal entities and state-linked actors from China, North Korea, and Russia are increasingly utilizing commercial AI tools to enhance the velocity and scale of their offensive capabilities.

該行動針對的是一個網頁版系統管理工具，利用語義邏輯缺陷來規避雙重認證。GTIG 透過 Python 腳本中出現的「幻覺」CVSS 分數與教科書式格式（這是 LLM 訓練數據的特徵）識別出人工智能的使用。雖然具體使用的模型尚未確定，但 Google 指出其 Gemini 模型可能未被利用。此事件與更廣泛的觀察一致，即來自中國、北韓與俄羅斯的犯罪實體與國家支持之行為者，正日益利用商業 AI 工具來提升其攻擊能力的速度與規模。

Concurrent with these developments, the emergence of highly capable models, such as Anthropic's Mythos, has necessitated a strategic shift in defensive postures. Anthropic restricted the release of Mythos due to its capacity to identify zero-day vulnerabilities across major operating systems and browsers, subsequently establishing Project Glasswing to coordinate security efforts among major technology and financial institutions. Similarly, OpenAI has introduced a specialized cybersecurity iteration of its model, restricted to vetted infrastructure defenders.

與此同時，如 Anthropic 的 Mythos 等高效能模型的出現，使得防禦態勢必須進行戰略轉移。由於 Mythos 具有識別各大操作系統與瀏覽器零日漏洞的能力，Anthropic 限制了其發佈，隨後成立 Project Glasswing 以協調大型科技與金融機構間的安全性工作。同樣地，OpenAI 推出了一個專為網絡安全設計的模型迭代版本，僅限於經過審核的基礎設施防禦者使用。

From a policy perspective, the United States administration has exhibited fluctuating stances regarding AI oversight. Despite an initial commitment to repeal previous regulatory guardrails, the Commerce Department recently entered agreements with Google, Microsoft, and xAI to evaluate high-capacity models prior to public dissemination, though the public record of these agreements was subsequently removed. Policy analysts suggest that while AI may eventually facilitate the hardening of legacy software, a transitional period of heightened systemic risk is anticipated as the capacity for automated exploitation currently outpaces the speed of defensive remediation.

從政策角度來看，美國政府在 AI 監管方面表現出波動的立場。儘管最初承諾廢除先前的監管限制，但商務部最近與 Google、Microsoft 及 xAI 達成協議，在高效能模型公開發佈前進行評估，儘管這些協議的公開紀錄隨後被移除。政策分析師認為，雖然 AI 最終可能有助於強化舊有軟體，但由於自動化利用的能力目前超過了防禦修復的速度，預計將進入一個系統性風險增加的過渡期。

The current landscape is characterized by an active race between AI-driven offensive exploitation and the development of coordinated institutional defenses.

目前的局面特徵在於 AI 驅動的攻擊利用與協調一致的機構防禦之間，正處於激烈的競爭狀態。