Back to skill

Security audit

feishu-multi-agent

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent multi-agent Feishu orchestration guide, but it explicitly tells agents to bypass workspace isolation and run persistent automated agents with broad cross-workspace access.

Review before installing. Use this only if you intentionally want a Feishu-connected OpenClaw setup with persistent cron jobs and an orchestrator that can reach managed agent workspaces through shell commands. Prefer a dedicated shared queue or broker with validation, producer identity, audit logs, and narrow permissions before using it in production, and define explicit consent and limits for proactive user-facing messages.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:147
Finding

Explicit Bypass of Workspace Isolation Through Unrestricted Shell Commands

Content
View full analysis
~/.openclaw/workspace-buyer/tasks/pending/003-xxx.md << 'EOF' # Task title Priority: Medium Description: ... EOF ``` Related instructions explicitly recommend using shell commands to bypass tool restrictions: ```markdown | Scenario | Solution | |----------|----------| | Foreman writes to another agent's workspace | Use bash `cat > file` instead of the Write tool | | Agent needs to read another workspace | Use bash `cat file` or `ls dir` | ``` ### Technical Analysis The skill recognizes that OpenClaw's sandboxed Read and Write tools enforce workspace boundaries through canonical path validation. It then directs agents to circumvent that control by invoking unrestricted shell commands. This changes the workspace boundary from an enforced security control into a convention. An agent with shell access can read or modify files in any location available to the operating-system account, not merely the intended task queue. The provided command demonstrates writes into another agent's workspace, while the related guidance also recommends cross-workspace reads. Although cross-agent task dispatch is part of the stated architecture, unrestricted shell access is broader than the legitimate requirement. It does not constrain access to a dedicated queue, validate task contents, authenticate task producers, or prevent access to other files in subordinate workspaces. ### Attack Path 1. An attacker sends malicious or prompt-injection content through a channel processed by the main agent, such as a Feishu ...[truncated 1657 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The title and operative instructions present the skill as Chinese-language only, and later examples require Chinese prompts and reports to users, with no indication that users can choose another language. This is a natural-language locale policy concern because the skill appears to impose a specific language without opt-in or justification.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 106)May include surrounding context.

bash
# 为每个受管 agent 创建任务目录
mkdir -p ~/.openclaw/workspace-buyer/tasks/{pending,done}
mkdir -p ~/.openclaw/workspace-kb/tasks/{pending,done}
mkdir -p ~/.openclaw/workspace-gf/tasks/{pending,done}

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

This duplicate finding points to the same behavior: persistent task-file injection into another agent's workspace via shell redirection. Because those files act as durable instructions, they can be used to influence later autonomous behavior without relying on an immediate interactive prompt.

Content

Scanner excerpt · SKILL.md (reported line 147)May include surrounding context.

md
- `ls ~/.openclaw/workspace-<id>/tasks/pending/` 检查任务队列

## ⚠️ 派活方式(重要)
必须用 bash 命令写任务文件,不要用 Write 工具!

    cat > ~/.openclaw/workspace-buyer/tasks/pending/003-xxx.md << 'EOF'
    # 任务标题

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

This duplicate finding points to the same behavior: persistent task-file injection into another agent's workspace via shell redirection. Because those files act as durable instructions, they can be used to influence later autonomous behavior without relying on an immediate interactive prompt.

Content

Scanner excerpt · SKILL.md (reported line 147)May include surrounding context.

md
- `ls ~/.openclaw/workspace-<id>/tasks/pending/` 检查任务队列

## ⚠️ 派活方式(重要)
必须用 bash 命令写任务文件,不要用 Write 工具!

    cat > ~/.openclaw/workspace-buyer/tasks/pending/003-xxx.md << 'EOF'
    # 任务标题

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill explicitly instructs the operator/agent to bypass OpenClaw's workspace sandbox by using shell commands to write directly into other agents' workspaces. This defeats an intended security boundary and enables unauthorized cross-workspace file access, which can lead to task injection, data tampering, and lateral movement between agents.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The heartbeat instructions direct an agent to proactively send random user-facing content, including '自拍' and casual interaction, when no task is pending. This creates a risk of unsolicited messaging, privacy-invasive or inappropriate content generation, and user trust issues because there is no consent, rate limiting, or policy boundary described.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

This section generalizes shell-based reads and writes across workspaces as the standard workaround for sandbox restrictions. Framing bypass behavior as an endorsed operational pattern normalizes evasion of isolation controls and increases the chance that agents can access or modify files outside their authorized boundary.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.