lov-do-they-love-me

Pass

Audited by Gen Agent Trust Hub on Oct 2, 2026

Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXPOSURE]: The skill processes WeChat message history (‘messages.jsonl’), which contains private conversations and potential PII. It employs a ‘PRIVATE’ regex in ‘scripts/matrix_dataset.py’ to filter out phone numbers, ID numbers, and addresses from generated quotes. While processing is local by default, ‘scripts/semantic_label.py’ allows for a remote API backend (‘--backend remote’), which would send chat text to an external provider. This feature is documented to require user consent and uses environment variables for API keys to prevent credential leakage.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted chat messages from the user’s history and interpolates them into an LLM prompt for topic classification in ‘scripts/semantic_label.py’. Maliciously crafted messages in the chat history could attempt to bypass the classification instructions, potentially leading to incorrect analysis results or biased relationship metrics.
  • [COMMAND_EXECUTION]: The workflow requires the execution of multiple Python scripts on the host system to process metrics, perform labeling, and generate SVG graphics. These scripts interact with the local file system and network (for LLM services), operating with the permissions of the user running the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 2, 2026, 05:24 PM