T09 · Insecure Skill Coding Practices
- Location
src/index.js:15- Finding
Unrestricted MCP URLs Permit Plaintext Transmission and Unsafe Network Destinations
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This appears to be a real MCP bridge, but it needs review because one approved tool can invoke many externally supplied tools without enough built-in limits.
Install only in trusted workspaces. Connect it only to MCP servers you control or trust, prefer HTTPS, avoid URL-embedded credentials, restrict the `mcp` tool to trusted/admin agents, and require separate review or confirmation before using database, account, messaging, filesystem, or other state-changing MCP tools. Do not share diagnostic reports until configuration and logs have been manually redacted.
src/index.js:15Unrestricted MCP URLs Permit Plaintext Transmission and Unsafe Network Destinations
src/index.js:129Single MCP Tool Bypasses Fine-Grained Authorization for Dynamically Discovered Capabilities
docs/TROUBLESHOOTING.md:410Diagnostic Report Workflow Collects Configuration and Logs for External Sharing Without Automatic Redaction
src/http-transport.js:32HTTP and SSE Responses Have No Time or Size Limits
Listing all MCP tools can reveal the names and capabilities of connected external services, which may expose sensitive operational context such as legal, database, or internal system integrations. In an agent environment, broad tool discovery without access controls can assist prompt-based reconnaissance and help an attacker identify higher-value tools to target next.
In OpenClaw chat:
User: List all MCP tools
AI: I'll check what MCP tools are available.
The testing example again encourages unrestricted enumeration of MCP tools from chat, reinforcing a workflow where tool inventory is exposed through conversational access. Even in a test scenario, this pattern is risky because production users often copy examples directly, potentially enabling reconnaissance against connected systems.
node test-server.js
User: List MCP tools
AI: [Shows available tools from test server]
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
**Option 3: .env file**
```bash
# ~/.openclaw/.env
KR_LEGAL_API_KEY=your-api-key-here
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
**Option 3: .env file**
```bash
# ~/.openclaw/.env
KR_LEGAL_API_KEY=your-api-key-here
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
{
"envFile": "/home/user/.openclaw/.env",
"plugins": { ... }
}
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
In OpenClaw chat:
User: List MCP tools
AI: Let me check available MCP tools.
[Uses: mcp with action=list]
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
In OpenClaw chat:
User: List MCP tools
AI: Let me check available MCP tools.
[Uses: mcp with action=list]
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
In OpenClaw chat:
User: List MCP tools
AI: Let me check available MCP tools.
[Uses: mcp with action=list]
Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
In OpenClaw chat:
User: List MCP tools
AI: Let me check available MCP tools.
[Uses: mcp with action=list]
This pattern attempts to override system instructions or ignore safety constraints. Without LLM analysis, manual review is recommended.
Diagnostic:
# Enable debug mode
# Edit http-transport.js:
# new StreamableHTTPClientTransport(url, { debug: true })
This pattern attempts to override system instructions or ignore safety constraints. Without LLM analysis, manual review is recommended.
Diagnostic:
# Enable debug mode
# Edit http-transport.js:
# new StreamableHTTPClientTransport(url, { debug: true })
The README explicitly demonstrates a natural-language workflow that causes an external MCP database tool to execute SQL queries, but it does not warn that such tools may reach real databases or require confirmation for destructive or sensitive operations. In an agent skill, normalizing direct database access through external tool bridges increases the chance of unsafe autonomous query execution, data exfiltration, or accidental modification if the connected MCP server exposes powerful operations.
The skill explicitly enables discovery and execution of external MCP tools across connected servers, which is a network-capable, potentially high-privilege integration surface, yet it declares no explicit tool scope or allowed-tools restrictions. This creates an overly broad trust boundary where an agent may invoke unexpected third-party tools or data sources without clear policy constraints, increasing the risk of data exfiltration, unintended actions, or unsafe tool use.
The front-matter description markets broad MCP connectivity to databases, APIs, legal systems, and other external services without warning that connected servers may expose sensitive, state-changing, or otherwise impactful operations. This omission can mislead users into treating all discovered tools as low-risk, when some may read confidential data or perform consequential actions.
The skill instructs the agent to discover and call external MCP tools but does not warn users that their prompts, parameters, or retrieved data may be transmitted to third-party MCP servers. Without this disclosure, users may unknowingly expose sensitive information to external services outside the primary trust boundary.
This is a manifest JSON file, so vague-trigger review applies. The description says the plugin can 'connect to MCP servers and use their tools' but provides no specificity about when it should be invoked, what contexts it applies to, or any limiting conditions, making the activation scope overly broad for a skill manifest.
The documentation advertises a single agent-exposed tool that can enumerate and invoke arbitrary tools across multiple MCP servers, including examples like database querying, but it provides no safety guidance, permission model, or warning about sensitive/system-impacting operations. In an agent setting, this can normalize unsafe use of high-impact external capabilities and increase the chance of unintended data access, destructive actions, or prompt-driven misuse when the agent treats all discovered tools as equally safe to call.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
# Test URL is accessible
curl http://localhost:3000/mcp
# Test with JSON-RPC request
curl -X POST http://localhost:3000/mcp \
The examples normalize direct querying of orders and displaying customer-linked records without any privacy guidance, access-control caveat, or data-minimization note. In an MCP integration skill, examples strongly influence agent behavior, so this can lead users or downstream agents to retrieve and expose production data more broadly than necessary.
The workflow demonstrates combining internal database-derived event information with an external weather service, but omits any warning that internal schedule/location data may be disclosed to third-party systems. Because the skill is specifically about centralized tool orchestration across MCP servers, this pattern increases the chance of unintended cross-system data sharing and privacy boundary violations.
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw logs | grep MCP
# Check plugin directory
ls -la ~/.openclaw/extensions/mcp-integration/
# Test configuration syntax
cat ~/.openclaw/openclaw.json | jq '.plugins.entries["mcp-integration"]'
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw logs | grep MCP
# Check plugin directory
ls -la ~/.openclaw/extensions/mcp-integration/
# Test configuration syntax
cat ~/.openclaw/openclaw.json | jq '.plugins.entries["mcp-integration"]'
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw logs | grep MCP
# Check plugin directory
ls -la ~/.openclaw/extensions/mcp-integration/
# Test configuration syntax
cat ~/.openclaw/openclaw.json | jq '.plugins.entries["mcp-integration"]'
Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.
openclaw logs | grep MCP
# Check plugin directory
ls -la ~/.openclaw/extensions/mcp-integration/
# Test configuration syntax
cat ~/.openclaw/openclaw.json | jq '.plugins.entries["mcp-integration"]'
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
netstat -tuln | grep 3000
curl -X POST http://localhost:3000/mcp
-H "Content-Type: application/json"
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'