T02 · Agent Memory Poisoning
- Location
SKILL.md:51- Finding
Untrusted conversation content can poison persistent Agent policy and identity files
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill has a coherent daily-retrospective purpose, but it asks for broad session-history access, persistent agent memory/config edits, scheduled execution, and ships an enabled real Feishu webhook.
Review before installing. Do not use the packaged Feishu webhook; revoke or replace it with your own secret and keep notifications disabled until you know exactly what will be sent. Run the skill manually or as an unprivileged user, not from root cron. Require a diff and explicit approval before any change to MEMORY.md, USER.md, SOUL.md, or AGENTS.md, and avoid storing raw conversation details, tool results, secrets, or sensitive profile inferences.
SKILL.md:51Untrusted conversation content can poison persistent Agent policy and identity files
INSTALL.md:88Daily cron registration creates privileged cross-session persistence
config.json:39Enabled Feishu webhook credential is committed in plaintext
agent_retro.py:99Retrospective workflow ingests complete session records without data minimization
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw --version
ls -la ~/.openclaw/workspace/
## 🚀 安装步骤
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw --version
ls -la ~/.openclaw/workspace/
## 🚀 安装步骤
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw --version
ls -la ~/.openclaw/workspace/
## 🚀 安装步骤
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
# 确保有写入权限
chmod 755 ~/.openclaw/workspace/memory
chmod 644 ~/.openclaw/workspace/skills/agent-retro/*.py
# 设置执行权限
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
# 确保有写入权限
chmod 755 ~/.openclaw/workspace/memory
chmod 644 ~/.openclaw/workspace/skills/agent-retro/*.py
# 设置执行权限
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
# 确保有写入权限
chmod 755 ~/.openclaw/workspace/memory
chmod 644 ~/.openclaw/workspace/skills/agent-retro/*.py
# 设置执行权限
chmod +x ~/.openclaw/workspace/skills/agent-retro/agent_retro.py
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
# 编辑crontab
crontab -e
# 添加以下行
5 0 * * * cd /root/.openclaw/workspace/skills/agent-retro && python3 agent_retro.py >> /var/log/agent_retro.log 2>&1
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
# 查看cron任务
crontab -l | grep agent_retro
# 查看cron日志
grep CRON /var/log/syslog | tail -10
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
# 查看cron任务
crontab -l | grep agent_retro
# 查看cron日志
grep CRON /var/log/syslog | tail -10
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
# 查看cron任务
crontab -l | grep agent_retro
# 查看cron日志
grep CRON /var/log/syslog | tail -10
The maintenance section recommends automatic deletion of backup and lock files using find ... -delete without warning that the action is irreversible. If paths are misconfigured or users do not understand what matches the pattern, important recovery data or state files may be permanently removed.
The upgrade instructions include a destructive command (rm -rf agent-retro) without an explicit warning, confirmation step, or safer alternative. In an installation guide, this can cause irreversible deletion of the existing skill directory or unintended data loss if the user is in the wrong path or the command is modified incorrectly.
The document presents the skill entirely in Chinese and names the system in Chinese, with no indication that language is configurable or that the skill is intentionally limited to a Chinese-speaking context. Under the policy, forcing a specific language without user opt-in can be a natural-language policy violation unless the locale constraint is explicitly documented and justified.
The README describes an automated system that analyzes prior conversations, updates multiple persistent memory/configuration files, and sends summaries to Feishu, but it does not present any explicit privacy warning, consent requirement, or data-handling boundaries. Because the processed content appears to include conversation-derived user information and behavioral profiles, this creates a real risk of silent collection, retention, and third-party disclosure of sensitive data.
The README instructs the skill to analyze conversation history, derive user and agent profiles, write that information into persistent files, and send reports through Feishu. This is a substantive data-handling vulnerability because it operationalizes extraction and redistribution of conversation-derived personal or sensitive information without visible minimization, access control, consent, or sanitization safeguards.
The skill instructs the agent to read all session records for the target date and extract user commands, assistant replies, tool calls, and tool results without any privacy notice, minimization rule, or consent check. This creates unnecessary exposure of potentially sensitive history and can propagate private data into derived memory and profile files, increasing confidentiality risk.
The skill directs the agent to perform broad, persistent modifications across multiple workspace files, including core configuration files, and explicitly requires real physical writes before reporting to the user. This is dangerous because it can silently alter long-lived state, overwrite trusted context, and create hard-to-detect persistence or integrity issues without explicit user consent at execution time.
The module description and all user-facing strings are written exclusively in Chinese, indicating the skill is designed to operate in a fixed language without offering a user choice. This can violate language/locale policy when no opt-in or justified region-specific constraint is documented.
The workflow analyzes session content and persists derived summaries such as user profile, agent profile, actions, and mistakes into long-term markdown files in plain language. In this context, the danger is elevated because the code fabricates or infers sensitive profile traits and stores them unencrypted, which can expose personal data, operational habits, and conversation-derived inferences to other local users, backups, or later tooling.
The skill writes retrospective summaries, user profile data, and agent profile data to persistent files under the workspace without explicit consent, confirmation, or a prominent warning. Because the content is derived from conversations, this can silently store sensitive personal or operational information in long-term memory, increasing privacy and data-retention risk.
The function name, comments, and caller-facing logs indicate that core configuration files are being updated, but the implementation only creates backups and returns a list of backup actions. This is an integrity and trust issue: operators may believe state was changed when it was not, which can cause security assumptions, audit errors, or skipped follow-up actions.
The generated report states that core configuration files were updated and that safety measures were applied, but the code path only backs up existing files. Misleading security/status reporting is dangerous because users may rely on false evidence of completed hardening or persistence actions.
This is a manifest file, so vague-trigger review applies. The description "Agent每日复盘系统 - 基于Tangc规范实现" describes the skill's purpose in broad terms but provides no specific trigger phrases, scope limits, or exclusion conditions, which can make invocation criteria unclear in systems that infer activation from manifest text.
The visible natural-language metadata is entirely in Chinese (for example, the description and author fields), and there is no indication that users can choose another language or opt in to a locale-specific experience. Under the policy, forcing a specific language without user choice is a natural-language policy concern unless the locale restriction is explicitly documented and justified.
The configuration embeds a live Feishu bot webhook and enables outbound messaging by default, creating an exfiltration and unsolicited notification channel. In an agent skill that processes session history, memory files, and agent profiles, this can transmit sensitive operational or user-derived data to an external service without clear scoping, consent, or justification beyond a generic output channel.