T09 · Insecure Skill Coding Practices
- Location
analyzer.py:75- Finding
Unrestricted Path Traversal Through Client and Session Arguments
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill has a coherent livestream-report purpose, but it handles sensitive customer materials with under-scoped file paths, broad prompt logging, and an optional external model call that can send data and an API key to a configurable endpoint.
Review this skill carefully before installing. Use it only with data you are allowed to send to the configured model provider, keep ~/.openclaw/config.json protected, avoid custom or non-HTTPS review_api_url values, and do not pass untrusted --client or --session values until path validation is added. Treat stdout and generated prompt files as sensitive because they can contain the full customer profile, livestream data, and transcript.
analyzer.py:75Unrestricted Path Traversal Through Client and Session Arguments
analyzer.py:102Sensitive Customer Materials Are Unconditionally Printed to Standard Output
analyzer.py:21API Credentials and Customer Content Can Be Sent to an Arbitrary Configured Endpoint
SKILL.md:8Unpinned and Unnecessary Third-Party Dependencies
The skill advertises behavior that reads local files, writes report outputs, and can send data over the network, but it does not declare any explicit tool scope or permission boundaries. This increases the chance that an agent executes the skill with broader-than-necessary capabilities, reducing transparency and weakening least-privilege controls.
The documentation says --call-model will directly call an external API to generate the report, but it does not clearly warn that live-stream data, user profiles, and scripts may be transmitted to a third-party endpoint. Because these inputs can contain sensitive business or personal information, users may unknowingly exfiltrate confidential data to an external service.
The file’s natural-language strings, descriptions, and interface text are entirely in Chinese, suggesting the skill is designed around a fixed language/locale. There is no indication in this file that users may opt into another language or that the Chinese-only behavior is a documented region-specific requirement.
The code sends loaded session data, user profile data, and transcript content to a configurable external endpoint via review_api_url without any explicit consent gate, destination allowlisting, or user-facing disclosure at the point of transfer. Because the endpoint is read from local config and defaults to a third-party URL, sensitive customer/business data can be exfiltrated to an unexpected service, making the skill more dangerous in this context since it processes replay and growth-planning material that may contain proprietary or personal information.
The prompt instructs the model entirely in Chinese and requires a specific Chinese-language reporting style ('你的语言要通俗易懂'). This imposes a fixed language/locale behavior without any user opt-in or explanation that the skill is intended only for Chinese-language users or a region-specific workflow.
Natural-language policy issues include forcing a specific language without user opt-in. This file presents the skill description, parameters, and operational guidance only in Chinese and does not state that the skill is region-specific or offer an alternative language option.
The manifest description states the skill will automatically generate a detailed report based on customer and session inputs. However, the implementation makes model invocation optional via --call-model; without that flag, it only assembles and prints the prompt text rather than generating the promised report. This is a mild behavior mismatch between the declared outcome and the actual default behavior.