重點導讀
導語:Check Point 發現 Codex jailbreak,能透過簡單「hi」指令讓 LLM 代理自動回覆廣告,目標為 zz.aiapi2025.top。
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:
Thomas Roccia 🤘@fr0gger_
·16 天前
🧐 Prompt persistence for ads!
Check Point found a Codex jailbreak that rewrites the global instructions of an agent and add a Skills so that a simple "hi" makes the agent reply with an ad for an LLM gateway (zz.aiapi2025[.]top).
Few techniques used here 👇
- ad injected in 00-identity.md, plus six seagull-* Skills
- triggers: hi / hello / ctf / coach / 海鷗
- payload.pkg is AES-256-GCM encrypted and Ed25519-signed
Hunting opportunity:
- unexpected writes to AGENTS.md, CLAUDE.md, Skills dirs
- zz.aiapi2025[.]top, 海鷗線上, 整點薯條
- seagull-* skill folders
Prompt available in PromptIntel including a NOVA rule.
推文來源 · X (Twitter)在 X 參與討論
來源: Thomas Roccia 🤘 (x.com)
在 X 查看推文