The Agent Skills Directory

[SAFE]: The skill is designed with a strong security posture, prioritizing data privacy and offline operation. It implements a 'fail-closed' mechanism where any input not matching the strict JSON schema is rejected.
[DATA_EXFILTRATION]: The skill prevents data exposure by operating exclusively on HMAC-SHA256 pseudonyms generated outside the skill environment. The analysis script lacks network capabilities and recursively scans input for forbidden keywords (e.g., 'email', 'secret', 'token') to prevent the ingestion of sensitive identity data.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests external JSON data, the risk of indirect injection is mitigated by strict type checking, exact field matching, and the absence of any tool or command execution based on data content.
[PROMPT_INJECTION]: Static analysis identified injection patterns in eval-spec.yaml; however, these are contained within the test suite's adversarial test cases. This represents a security best practice—testing the agent's ability to resist instructions like 'Ignore previous instructions'—rather than a vulnerability.
[COMMAND_EXECUTION]: The skill uses the Bash tool solely to execute a local, self-contained Python script. This script follows secure coding practices, avoiding dynamic execution functions like eval() or exec() and performing only deterministic logic on validated data structures.

grammarly-license-governor