The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The scripts/lm_cookie.py utility retrieves session cookies from the user's browser using the Chrome DevTools Protocol (CDP) on port 9222. It leverages the Network.getCookies method to access authentication tokens for targeted legal platforms. This functionality enables the agent to harvest sensitive credentials from an active browser session without user-provided passwords.
[DYNAMIC_CONTEXT_INJECTION]: SKILL.md contains a shell command execution (! ... ) that runs during the skill's initialization phase. The command attempts to find and execute the skill-toc utility to generate a table of contents, representing the execution of dynamic shell logic at load time.
[COMMAND_EXECUTION]: The scripts/harvest_complaints.py script uses subprocess.run to invoke system binaries such as pdfinfo and pdftotext. These calls process file paths constructed from external parquet metadata, creating a potential surface for command injection if the input files are compromised.
[DYNAMIC_EXECUTION]: Several scripts (da_grab.py, fetch_opinion_pdfs.py, harvest_complaints.py) perform dynamic module loading by searching parent directories for a ds/scripts path and appending it to sys.path. This dynamic path computation for internal dependencies introduces risks associated with runtime module resolution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from court dockets and PDF filings, creating a surface for indirect prompt injection.
Ingestion points: External data enters through da_docket_metadata.parquet, chancery_case_panel.parquet, and downloaded legal document PDFs.
Boundary markers: The skill lacks explicit security delimiters or sanitization routines to isolate document text from agent instructions.
Capability inventory: The skill possesses network access, browser interaction capabilities via websockets, system command execution, and file system operations.
Sanitization: No robust sanitization is implemented to prevent instructions embedded within legal filings from affecting the agent's logic.

court-dockets