T02 · Agent Memory Poisoning
- Location
dream/dream_cycle.py:58- Finding
Persistent agent memory poisoning through untrusted model output
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is purpose-aligned as a continuous-learning assistant, but it creates persistent jobs, exports stored memories to an external model API, and lets model output modify future agent-governing documents without enough user control.
Review before installing. Only use this skill if you are comfortable with persistent local memory, unattended daily jobs, and possible transfer of stored conversations and metadata to a configured model provider. Do not store secrets or regulated data in MemPalace with this enabled, and avoid installing the scheduler until there is explicit opt-in, redaction, review-before-write, and an uninstall path.
dream/dream_cycle.py:58Persistent agent memory poisoning through untrusted model output
dream/dream_cycle.py:177Bulk disclosure of persistent memories and metadata to an external API
setup/install_cron.py:38Persistent daily execution through Scheduled Tasks and cron
setup/install_cron.py:104Cron command injection through unquoted workspace and script paths
setup/init_learning_files.py:17Initialization destructively overwrites existing persistent learning files
config/dream_config.json:18Runtime configuration schema does not match the documented security settings
The skill claims a broad autonomous learning workflow, but the described concrete actions emphasize local config writes, notification queues, and sync/notification behavior rather than clearly bounded implementations of the advertised features. This mismatch can mislead users about what data is collected, what files are changed, and what automation actually occurs, reducing informed consent and auditability.
The skill claims a broad autonomous learning workflow, but the described concrete actions emphasize local config writes, notification queues, and sync/notification behavior rather than clearly bounded implementations of the advertised features. This mismatch can mislead users about what data is collected, what files are changed, and what automation actually occurs, reducing informed consent and auditability.
The skill describes automatic recording of conversations and scheduled analysis, but it does not present a clear upfront warning that user conversations and preferences will be persisted and processed later. This is dangerous because users may disclose sensitive information without understanding that it will be stored across sessions and mined into long-lived profile documents.
Referenced artifact was not completely inspected
python setup/install_cron.py
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 skill posts diary contents and metadata to a remote API without any visible user warning, consent gate, or privacy control. Because the source data is effectively a memory store, this can leak highly sensitive information off-device and create compliance and confidentiality issues.
Using shell=True for schtasks is a tool-parameter misuse because it broadens execution semantics beyond the intended program invocation. In an adversarial or polluted environment, shell lookup and parsing can be abused to run the wrong binary or otherwise alter behavior.
def get_existing_windows_tasks():
"""获取Windows现有任务"""
try:
result = subprocess.run(
['schtasks', '/query', '/fo', 'LIST', '/v'],
capture_output=True,
text=True,
This is another instance of unsafe subprocess parameterization: task deletion is performed with shell=True despite no need for shell features. If execution is hijacked, the installer could perform arbitrary actions while appearing to manage scheduled tasks.
# 删除旧任务
for task_name in ['OpenClaw-ContinuousSync', 'OpenClaw-ContinuousDream']:
try:
subprocess.run(
['schtasks', '/delete', '/tn', task_name, '/f'],
capture_output=True,
shell=True
This task creation call is especially sensitive because it establishes recurring execution. The combination of shell=True and a scheduler action string containing a path derived from the current workspace creates an avoidable avenue for command/parameter abuse and persistent unintended execution.
success = True
for task in tasks:
try:
result = subprocess.run(
[
'schtasks', '/create',
'/tn', task['name'],
The README explicitly promotes automatic archival of daily conversations, semantic memory storage, and cross-session context retention, but it does not warn users about privacy implications, retention duration, consent, or handling of sensitive data. In a continuous-learning skill, this omission is dangerous because users may unknowingly persist confidential prompts, credentials, personal data, or regulated information across sessions.
The workflow states that chat-derived content is analyzed on a schedule and then sent via WeChat notification, but the README does not clearly disclose that conversation content may be transmitted to external services or channels. This creates a real risk of unintended data exfiltration, especially because the skill also references external analysis services and automated unattended processing.
The skill documents capabilities that imply broad file I/O, shell execution, network access, and scheduled background execution, but it does not declare any explicit tool scope or permission boundaries. That makes the skill harder to audit and allows an agent runtime to grant more privilege than users would reasonably expect, increasing the chance of unintended file modification, persistence, or data exfiltration.
The trigger phrases are broad terms related to learning, memory, and self-improvement, which can easily match ordinary conversation. Overbroad activation increases the risk that the skill runs unexpectedly, causing unanticipated persistence of conversation data, file updates, or scheduled-task setup without clear user intent.
The manifest description and trigger phrases are entirely Chinese and imply operation through Chinese-language invocation, but the skill does not state that language is optional or region-specific. Under the policy, a skill should not force a specific language or locale without user opt-in or clear justification.
The skill directs automatic retention of conversation content into persistent memory across sessions and instructs the agent to read profile and memory files when a new task is inferred. This creates a durable cross-session dossier of user and workflow data that may exceed user expectations and can expose sensitive material if the memory store or local files are later accessed by other agents, users, or integrations.
The new-task detection criteria rely on vague signals such as topic changes, first-time keywords, or phrases like 'start doing...', which are insufficiently bounded. Ambiguous activation can cause the skill to read additional local memory files and policy documents unexpectedly, expanding data access based on heuristic guesses rather than clear user authorization.
The documented memory schema includes broad categories such as user preferences, project background, important decisions, and lessons learned. Such open-ended collection encourages storing sensitive or confidential conversational details without data minimization rules, increasing privacy and confidentiality risk.
The markdown explains that analyzed conversation history is extracted into local documentation files, but it does not clearly warn users that these files will be automatically updated. Automatic documentation writes can unintentionally store secrets, internal project details, or personal preferences in locations that are later synced, shared, or committed.
The dream-analysis workflow instructs the agent to read all memory fragments and consolidate them into persistent core documents about the user, workflow, tools, and bootstrap rules. Bulk consolidation amplifies privacy risk because scattered sensitive details become easier to retrieve, reuse, or leak once centralized in durable summary documents.
The skill is configured to send analysis data to an external API endpoint, which introduces a clear external transmission path for conversation-derived memory content. Without explicit data-handling disclosure, minimization, and user consent, sensitive user or project information could be transferred to a third party unexpectedly.
"analysis_model": {
"provider": "minimax",
"model": "MiniMax-M2.7",
"api_url": "https://api.minimax.chat/v1",
"api_key": "your_api_key_here"
}
}
The documented use of cron/Task Scheduler establishes session persistence through recurring background execution. Persistence is not inherently malicious, but in this skill context it increases risk because conversation storage, analysis, notification, and file modification can continue automatically after the user stops interacting.
Linux:
crontab -l | grep openclaw
整个技能文档以中文规定启动流程、记忆管理和操作规范,但没有说明这是面向特定中文用户群体的限定场景,也没有提供语言切换或用户选择机制。按规则,这类未获用户选择的单一语言约束属于自然语言层面的语言/地区策略风险。
The new-task detection criteria are overly broad and include ambiguous triggers like topic shifts, first-time keywords, and 'any uncertain situation'. In a continuous-learning skill, this can cause the agent to activate memory/bootstrap behaviors during ordinary conversation, leading to unnecessary document reads and unintended persistence of user content.
The instructions require ongoing recording, summarization, daily synchronization, and long-term storage of user conversations and memories in plain-language documents and external systems like MemPalace. This creates a meaningful privacy and data-governance risk because sensitive user content may be retained, reprocessed, and propagated without clear minimization, consent, redaction, retention limits, or access controls.
This JSON manifest contains user-facing natural-language descriptions exclusively in Chinese across the purpose fields. Under the stated policy, forcing a specific language without user opt-in or documented justification is a locale-policy violation applicable to all file types.