The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary data provided in the state field of the input JSON, which could contain instructions intended to influence the agent's behavior.
Ingestion points: The state parameter in the request JSON (defined in SKILL.md) accepts arbitrary data from users or external sources.
Boundary markers: There are no explicit delimiters or instructions provided to separate untrusted data from the agent's instructions within the state field.
Capability inventory: The skill executes the jev CLI tool and transmits data to the TypeSafe AI external API.
Sanitization: No specific validation, escaping, or sanitization steps are mentioned to verify the content of the state field before it is passed to the CLI.
[COMMAND_EXECUTION]: The skill invokes the jev CLI tool or swift run jev on the host system. While this is the intended mechanism for the skill's functionality, it involves executing local commands based on structured input.
[DATA_EXFILTRATION]: The skill transmits the state content to the TypeSafe AI service. While this is the primary purpose of the skill and is documented with a privacy warning, users should be cautious about including sensitive or private information in the state object as it is sent to an external third-party domain.

jev