Anthropic has disclosed that it disrupted a campaign where a Russian state-sponsored threat actor used its Claude AI to build an automated, iterative pipeline for evading security tools, as reported by The Hacker News.
The operation has been attributed by Anthropic to a cluster it designates GTG-20006, or "Generative Threat Group." The company said this activity aligns with the known actor APT29, also referred to as Midnight and Cozy Bear, which has been linked to Russia's foreign intelligence service.
The development marks a notable shift beyond simple AI-assisted code generation. Rather than a one-time code-writing exercise, the actors embedded the model into a continuous cycle. They used Claude to analyze security tool logs and forensic reports to understand why their malware was detected, with the AI pinpointing the specific code triggering alerts and proposing modifications to obfuscate those elements. The actors then implemented the changes, redeployed the malware, and fed new detection results back into the system.
This approach directly challenges signature-based detection. When attackers can systematically rewrite and redeploy malicious code through automation, the timeline from detection to successful evasion compresses dramatically—from weeks potentially down to hours.
For defenders, the incident underscores the strategic importance of behavioral analysis, anomaly detection, and zero-trust architectures that assume breach, rather than relying solely on static pattern recognition.
Anthropic's disruption and disclosure also highlights the growing role of AI platform providers in monitoring for and mitigating advanced, state-sponsored misuse—a precedent that arrives as the dual-use risks of powerful AI models transition from theoretical concerns to operational realities.
根據《The Hacker News》報道，Anthropic披露已瓦解一項行動，其中一名俄羅斯國家級威脅攻擊者利用其Claude人工智能建構了一個用於規避安全工具的自動化、迭代式流程。
Anthropic將此行動歸因於一個被編號為GTG-20006的攻擊集群，即「生成式威脅組織」。該公司表示，此活動與已知攻擊者APT29（亦被稱為Midnight和Cozy Bear）的行為一致，後者據信與俄羅斯對外情報局有關聯。
此發展標誌著超越簡單人工智能輔助代碼生成的重大轉變。攻擊者並非僅進行一次性編碼作業，而是將模型嵌入一個持續循環中。他們使用Claude分析安全工具日誌及取證報告，以了解其惡意軟件被偵測的原因，人工智能精確指出觸發警報的具體代碼，並提出修改建議以混淆相關元素。攻擊者隨後實施變更，重新部署惡意軟件，並將新的偵測結果反饋至系統。
此方法直接挑戰基於特徵碼偵測機制的基礎。當攻擊者能夠透過自動化系統性地重寫並重新部署惡意代碼時，從偵測到成功規避的時間線急劇壓縮——可能由數週縮短至數小時。
對防禦者而言，此事件凸顯了行為分析、異常檢測及假定已遭入侵的零信任架構的戰略重要性，而非僅依賴靜態模式識別。
Anthropic的瓦解行動及披露亦凸顯了人工智能平台供應商在監控及緩解國家級高級攻擊方面日益增長的角色——此先例出現之際，強大人工智能模型的雙重用途風險正從理論擔憂轉變為現實操作挑戰。