The Agent Skills Directory

[PROMPT_INJECTION]: No evidence of prompt injection or behavior override instructions was found. The instructions are focused on guiding the AI agent through a structured legal analysis framework.
[DATA_EXFILTRATION]: The skill processes local Word documents and creates archives in a local 'archive/' directory. There are no network calls or attempts to access sensitive system files like SSH keys or cloud credentials. All external URL references point to the author's official GitHub repository ('cat-xierluo/legal-skills'), which are legitimate vendor resources.
[REMOTE_CODE_EXECUTION]: No patterns of remote script execution (e.g., piping curl to bash) were found. The skill relies on local scripts provided within the package.
[COMMAND_EXECUTION]: A test script ('scripts/tests/test_runtime_regressions.py') uses 'subprocess.run' to verify CLI help outputs, which is a benign testing pattern. The PowerShell wrapper ('run_apply_review_plan.ps1') is used for environment staging and local execution for Windows compatibility.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests untrusted contract data from DOCX files. However, it implements security mitigations by using 'defusedxml' for XML parsing and focuses on structural data extraction (review-plan.json) rather than unrestricted text execution. This represents a low-risk surface common to document processing skills.
[EXTERNAL_DOWNLOADS]: The 'README.md' and 'requirements.txt' files list standard, well-known Python dependencies ('defusedxml', 'lxml'). No downloads from untrusted or risky domains were detected.
[OBFUSCATION]: No obfuscated code, Base64-encoded strings representing commands, or hidden Unicode characters were identified.

contract-copilot