create-cupid

Warn

Audited by Gen Agent Trust Hub on May 3, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill dynamically generates and writes new SKILL.md files to the .claude/skills/ directory at runtime. These generated files contain instructions and context for secondary agents, effectively allowing the skill to install and define new executable agent behaviors based on user input and processed external data.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted external sources, including WeChat/QQ logs, social media text, and EXIF metadata. This content is incorporated into the logic and context of the generated advisor agents.
  • Ingestion points: Files are read by tools/wechat_parser.py, tools/qq_parser.py, tools/social_parser.py, and tools/photo_analyzer.py.
  • Boundary markers: Safety instructions are included in the generated SKILL.md template (via tools/skill_writer.py), but no boundary markers or 'ignore embedded instructions' warnings are applied during the data parsing and extraction phase.
  • Capability inventory: Generated skills are user-invocable and rely on the logic in prompts/advisor_framework.md to analyze behavior and provide advice.
  • Sanitization: The Python parsing scripts lack explicit sanitization or filtering to prevent malicious instructions embedded within the imported chat logs or social media text from influencing the agent's behavior.
  • [DATA_EXPOSURE_AND_EXFILTRATION]: The skill collects and stores highly sensitive personal information, such as private chat histories and GPS location metadata extracted from photos. This data is aggregated in a predictable local directory structure (.claude/skills/{slug}/), which increases the potential risk of data exposure if the local environment is accessed.
  • [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The main skill (SKILL.md) makes extensive use of the Bash tool to invoke several local Python scripts (tools/*.py) for parsing data, managing sessions, and handling versions. This grants the skill the capability to execute shell commands and perform significant local file system operations.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 3, 2026, 07:28 PM
Security Audit — agent-trust-hub — create-cupid