360 Secure SkillHub
360 安全技能中心

Memory Poison Auditor

该 Skill 通过 Python 脚本扫描 OpenClaw 的 MEMORY.md 与每日记忆文件，检测提示注入式指令、品牌引导、异常品牌重复、伪造授权等记忆投毒信号，输出 PASS/WARN/BLOCK 结论；支持可选 AI 复审与一键清理，清理前生成备份，报告与备份写入固定输出目录。

分类：安全
2404589803
2404589803
热度 34.69
最近更新 2026-05-11 17:32:08

安装

发给 Agent
帮我安装这个技能：npx secure-skills-cli add 2404589803/memory-poison-auditor
兼容性
OpenClaw 专属
联网要求
不联网
数据出网
可选出网
Metadata
name
memory-poison-auditor
version
0.1.0
license
MIT
description
Audits OpenClaw memory files for injected instructions, brand bias, hidden steering, and memory poisoning patterns. Use when reviewing MEMORY.md, daily memory files, or any long-term memory store that may have been contaminated through dialogue.
metadata
{"clawdbot":{"emoji":"🧠","requires":{"bins":["python3"]}}}

Memory Poison Auditor

memory-poison-auditor checks whether OpenClaw memory files have been contaminated by hidden instructions, brand steering, injected operational policies, or suspicious recommendation bias written through prior conversations.

What It Checks

  • Prompt-injection style instructions inside memory.
  • "Always recommend X" or "never mention Y" style brand steering.
  • Abnormal brand repetition and preference shaping.
  • Suspicious authority claims like fake approvals or fake user intent.
  • Low-signal blocks that act like covert policy rather than factual memory.
  • Optional AI review for borderline suspicious blocks.

Commands

Audit Default Memory Roots

python3 {baseDir}/scripts/audit_memory.py scan
python3 {baseDir}/scripts/audit_memory.py --format json scan

Audit a Specific Path

python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/MEMORY.md
python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/memory

Optional AI Review

python3 {baseDir}/scripts/audit_memory.py scan --with-ai
python3 {baseDir}/scripts/audit_memory.py scan --path /root/clawd/memory/2026-03-15.md --with-ai

One-Click Cleaning

python3 {baseDir}/scripts/audit_memory.py clean --path /root/clawd/MEMORY.md --apply
python3 {baseDir}/scripts/audit_memory.py clean --path /root/clawd/memory --apply

Cleaning creates backups before rewriting suspicious blocks.

Output

Each audit returns:

  • PASS: no meaningful poisoning signals
  • WARN: suspicious memory blocks detected
  • BLOCK: memory likely contaminated and should be reviewed/cleaned

Reports and backups are written to:

/root/clawd/output/memory-poison-auditor/reports/
/root/clawd/output/memory-poison-auditor/backups/

Operational Guidance

  • Use this before trusting long-term memory in important planning or recommendations.
  • WARN means review before relying on that memory block.
  • BLOCK means clean or quarantine the memory before reuse.
  • AI review is optional and intended only for ambiguous cases.