T02 · Agent Memory Poisoning
- Location
dream/dream_cycle.py:251- Finding
Persistent Agent Instruction Poisoning Through Untrusted Model Output
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed continuous-memory workflow, but it can automatically turn stored conversations into future agent instructions and send memory excerpts to an external model provider without strong user approval controls.
Review this skill before installing. Only use it in a workspace where you are comfortable with conversation history becoming persistent memory and possible future agent instructions. Disable or empty the external analysis API key until you explicitly choose a provider, avoid storing secrets in MemPalace, and manually review any proposed changes to AGENTS.md, BOOTSTRAP.md, SOUL.md, MEMORY.md, or TOOLS.md before letting future agents rely on them.
dream/dream_cycle.py:251Persistent Agent Instruction Poisoning Through Untrusted Model Output
dream/dream_cycle.py:168Sensitive Memory Content Is Sent to a Configurable External API Without Redaction or Safe Opt-In
setup/init_learning_files.py:16Initialization Script Silently Overwrites Existing Learning Records
requirements.txt:4Unbounded Dependency Versions Create Reproducibility and Supply-Chain Risk
The documented workflow instructs the system to sync chats, analyze them, extract distilled content, and send notifications, creating a direct path for user conversation data to be transformed and transmitted beyond the original interaction boundary. This is especially risky because summaries can still contain sensitive information, and scheduled automation reduces user visibility and control.
This second mismatch finding highlights that the skill claims autonomous learning features while also performing message sending and notification-queue writes, yet several core advertised behaviors are not evidenced in the described implementation. Such gaps can mislead users into granting trust to a workflow that handles conversation-derived data and external signaling differently than expected.
This second mismatch finding highlights that the skill claims autonomous learning features while also performing message sending and notification-queue writes, yet several core advertised behaviors are not evidenced in the described implementation. Such gaps can mislead users into granting trust to a workflow that handles conversation-derived data and external signaling differently than expected.
The skill describes automatic conversation recording and long-term memory consolidation but does not present a clear, upfront warning or consent mechanism for this data collection. This is dangerous because users may disclose sensitive personal, project, or operational information without understanding it will be retained across sessions and summarized into persistent documents.
The skill instructs automatic retention and summarization of user conversations into persistent memory across sessions. This is dangerous because it creates long-lived profiles and records without clear minimization boundaries, potentially capturing secrets, personal data, business context, and mistakes that can later be exposed, misused, or transmitted.
The memory record format explicitly stores conversation summaries, user preferences, project background, important decisions, and error lessons. Consolidating this breadth of sensitive context in a persistent store materially raises privacy, profiling, and data leakage risks, especially when paired with external APIs and notifications.
The dream-analysis workflow reads all memory fragments and promotes them into long-term profile and rule documents such as SOUL.md, AGENTS.md, MEMORY.md, and TOOLS.md. This kind of automated consolidation can amplify harm by turning transient conversation data into durable behavioral rules and personal profiles, potentially including inaccurate or overly sensitive inferences.
YARA rule matched a known malware signature (reverse shell, backdoor, ransomware, C2 framework, or info stealer).
5 } }
### 分析模型配置
支持多个模型:
- **MiniMax M2.7**(默认,效果最好)
- **智谱GLM-4.7**
- **Claude 3.5 Sonnet**
- **GPT-4o**
---
## 🔧 故障排查
### 问题1:定时任务不执行
**Windows**:
```powershell
schtasks /query /tn "OpenClaw-SyncWeChat"
schtasks /query /tn "OpenClaw-DreamCycle"
检查状态是否为"就绪"。
Linux:
crontab -l | grep openclaw
检查:
检查:
添加自己的文档:
{
"PROJECT_NOTES": {
"path": "docs/project_notes.md",
"purpose": "项目特定笔记"
},
"API_REFERENCE": {
"path": "docs/api_reference.md",
"purpose": "API调用记录"
}
}
``
The README explicitly advertises automatic archiving of daily conversations, semantic memory storage, and cross-session context retention, but it provides no clear privacy notice, consent flow, retention limits, or guidance on handling sensitive data. In a memory-management skill, this can cause users to unknowingly persist personal, confidential, or regulated information across sessions.
The skill is designed to retain and reuse conversation content across sessions, which creates a real security and privacy risk if sensitive user information is stored without strict controls. Even if intended as a convenience feature, persistent memory broadens exposure in the event of unauthorized access, misuse, or accidental retrieval in later conversations.
The README describes scheduled background analysis at 02:00 and automated notifications, but does not clearly warn users that chat-derived content may be processed unattended and transmitted externally. Background processing increases the risk of silent handling of sensitive data, especially if summaries or extracted insights are pushed to notification channels such as WeChat.
The configuration instructions tell users to add a MiniMax API key for analysis but do not warn that conversation-derived data may be sent to a third-party model provider. This omission is dangerous because users may assume processing is local while in reality stored memories, chat summaries, or extracted insights could be transmitted off-device to an external service.
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.
Linux:
crontab -l
The skill describes capabilities that imply file reads/writes, shell execution, scheduling, and network access, but it does not declare any explicit permission or allowed-tool scope. This creates an overbroad trust boundary where an agent may invoke sensitive capabilities without clear user visibility or restriction, increasing the chance of unintended data access, persistence, or external transmission.
Broad trigger phrases such as '持续学习' or '记忆管理' are common conversational terms and could activate the skill unintentionally during ordinary discussion. In a skill that records conversations, updates memory, and may trigger scheduled or external actions, accidental activation materially increases privacy and integrity risk.
The primary description, examples, and trigger phrases are presented entirely in Chinese, and the file does not state whether other languages are supported or whether Chinese is an intentional region-specific requirement. This can amount to a locale policy issue when a skill effectively forces one language without user opt-in or justification.
The skill mentions WeChat notifications tied to synchronization and dream-analysis events but does not clearly warn that conversation-derived activity metadata may be sent to an external channel. Even if message contents are brief, outbound notifications can leak timing, task existence, or sensitive workflow context to third parties or shared accounts.
The skill is configured to send data to an external API endpoint for analysis. External transmission is not inherently malicious, but in this skill's context it is risky because the transmitted content may include memory fragments, user preferences, project context, or conversation summaries derived from persistent storage.
"analysis_model": {
"provider": "minimax",
"model": "MiniMax-M2.7",
"api_url": "https://api.minimax.chat/v1",
"api_key": "your_api_key_here"
}
}
Remembering a user's preferred form of address across future conversations is less severe than storing full project history, but it still represents cross-session personal profiling and persistent preference storage. Without explicit consent and visibility, even seemingly harmless preferences can normalize unauthorized long-term retention.
The skill includes instructions and mechanisms for scheduled execution and session persistence via cron/task scheduling. Persistence is not automatically malicious, but in a skill that stores conversation-derived data and performs outbound actions, unattended recurring execution increases the blast radius of mistakes, silent data accumulation, and unnoticed external signaling.
Linux:
crontab -l | grep openclaw
The skill file is written as a Chinese-only operating rule set and does not indicate that the user may choose another language or locale. This can constitute a language-policy issue when the skill behavior is constrained to a specific language without explicit opt-in or documented regional justification.
The new-task detector uses very broad criteria such as '关键词第一次出现' and '任何不确定的情况', which can cause the agent to misclassify ordinary conversation turns as new tasks. In this skill, that misclassification is security-relevant because a 'new task' triggers reading multiple memory and rule files, increasing unnecessary access to potentially sensitive context and making prompt-routing easier to manipulate.
The skill explicitly instructs the agent to record conversation content into long-term memory, sync chats on a schedule, and send notifications, which creates a concrete risk of storing and transmitting sensitive user data without contextual consent or minimization. In a continuous-learning skill, this is more dangerous because persistence and periodic exfiltration are core workflow features, so a single sensitive conversation could be retained and redistributed automatically.
This JSON manifest uses Chinese-only natural-language purpose descriptions across all documentation targets. Because the file provides no indication that the skill is region-specific or that users can choose their preferred language, it may violate a language/locale policy requiring opt-in or justification for enforced language constraints.
The configuration sends analysis data to an external API endpoint, which creates a real data egress path from the learning/memory workflow to a third-party service. In the context of a continuous-learning skill that may process conversation logs and memory artifacts, this increases the risk of leaking sensitive user data, especially if data minimization, consent, and redaction are not enforced.
"analysis": {
"model": "minimax",
"model_name": "MiniMax-M2.7",
"api_url": "https://api.minimax.chat/v1",
"api_key": "your_minimax_api_key_here",
"timeout_seconds": 120,
"temperature": 0.3