許多 AI 攻擊框架會產生大量請求,容易被 SOC 警報偵測。
- 作者指出 Storm-3358 威脅團隊在使用 Hermes 與 ARTEX 時,指示代理人改用只讀 robot 帳號,且將「58 次 / 7 分鐘」的存取速率降至「10 次 / 30 分鐘」,以降低噪音。
- 這種減速與許可權限制的手法讓惡意行為更不易被發現,提醒防禦者在偵測規則上需要考慮低頻、低許可權的異常行為。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
🧐 Storm-3358 threat actor is also using Hermes in addition to ARTEX framework to compromise Orgs.
In one of the instruction files, they teach the agent to avoid actions that could trigger alerts in your SOC and instruct the agent to do differently and to adapt accordingly.
For example, in a skill called redteam-detection-surface-and-resilience they put:
"資料面讀取用只讀 robot 賬號,不要用 admin;並且控速(把「58 次 / 7 分鐘」降成「10 次 / 30 分鐘」)"
Translated:
"When reading data, use a read only robot account, not admin, and slow down: go from '58 requests in 7 minutes' to '10 requests in 30 minutes'."
Interesting to see that AI offensive framework are noisy so they find a way to reduce this.
File available in PromptIntel.
IOC: 199[.]19[.]95[.]68