The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external CRM systems and webhooks, representing a typical surface for indirect instruction injection.
Ingestion points: Lead data is received via the /score endpoint and CRM data synchronization routines.
Boundary markers: No explicit delimiters are used to separate untrusted data from processing logic in the current code.
Capability inventory: The skill can perform network requests to external APIs and read/write to the local filesystem for model management.
Sanitization: Relies on Pydantic for data type validation, which does not filter for malicious instruction patterns in string fields.
[DYNAMIC_EXECUTION]: The skill utilizes serialization libraries to load pre-trained models, which involves executing logic stored in the model files.
Evidence: The application uses joblib and a configurable MODEL_PATH to load .pkl files.
Risk Factor: While standard for ML, loading untrusted or tampered serialization files can lead to arbitrary code execution.

ai-lead-scoring