The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it ingests untrusted data from external sources to build user voice profiles.
Ingestion points: User-provided text samples and LinkedIn activity (posts/comments) pulled via external connectors like Handlus or Apify (documented in sub-skills/voice-profile.md).
Boundary markers: No specific delimiters or instructions to ignore embedded commands are implemented for the ingested data.
Capability inventory: The skill can execute shell scripts (scripts/test_detectors.py), perform network operations to various third-party APIs, and write to local files (references/voice-profile.md).
Sanitization: No explicit sanitization or validation of the ingested LinkedIn content is described.
[DATA_EXFILTRATION]: The skill intentionally sends user-provided text to multiple external AI detection services (GPTZero, Originality.ai, ZeroGPT, Sapling, Copyleaks) to compare their results.
Evidence: sub-skills/detector-tester.md explicitly states that text is uploaded to these providers and warns users not to run it on confidential information.
[COMMAND_EXECUTION]: The skill instructions rely on the execution of local Python scripts to perform detection tests and configuration checks.
Evidence: Multiple references to python3 scripts/test_detectors.py and python3 scripts/check_config.py are found in SKILL.md and sub-skills/detector-tester.md.

linkedin-humanizer