The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a pattern for handling ContentViolationError (422 status code) where the agent is instructed to retry a generation using a suggested_prompt provided within the API error response body. This creates an attack surface where an external source (the API response) can influence the agent's future instructions.
Ingestion points: The suggested_prompt field in the JSON body of a 422 ContentViolationError response, as described in SKILL.md.
Boundary markers: The instructions provide a mitigation by stating the agent should only retry with the suggested prompt "if the user consents," establishing a human-in-the-loop verification step.
Capability inventory: The reference implementation (callVenice) and surrounding instructions involve performing and retrying network-based generation requests with modified parameters.
Sanitization: No specific automated sanitization or filtering of the suggested_prompt content is mentioned.

venice-errors