The Agent Skills Directory

[Indirect Prompt Injection] (HIGH): The documentation describes features that ingest external data (JSON, Python dicts) and map them to types with execution side effects. Ingestion points: model_validate_json (references/examples/files.md), TypeAdapter.validate_python (references/concepts/type_adapter.md). Boundary markers: Absent in example code. Capability inventory: ImportString (references/api/types/ImportString.md), create_model (references/api/base_model/create_model.md). Sanitization: Not demonstrated; documentation explicitly warns about RCE in create_model.
[Remote Code Execution] (HIGH): The create_model function and ImportString type provide mechanisms for dynamic code evaluation and module loading. The documentation for create_model.md explicitly warns that it can execute arbitrary code contained in field annotations. ImportString.md demonstrates importing Python objects from strings, which is exploitable if inputs are attacker-controlled.
[Dynamic Execution] (MEDIUM): Multiple files describe runtime code generation or dynamic loading. ImportString performs dynamic loading from computed paths. create_model performs dynamic code generation based on annotations. These are documented as features but represent high-risk primitives in an AI agent context.

pydantic-dev