The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from local and remote API specifications without implementing boundary markers or explicit sanitization.
Ingestion points: The skill searches for and reads local files matching **/openapi.{json,yaml,yml} and **/swagger.{json,yaml,yml} as described in references/sync-collections.md, and retrieves remote Postman collections and specs in references/api-documentation.md.
Boundary markers: The workflows do not include instructions for the agent to treat external specification content as data or to ignore embedded instructions.
Capability inventory: The agent has the capability to write new source code files to the local project and perform network operations through the Postman MCP Server.
Sanitization: There is no documented validation of the specification content before it is used for code generation or documentation updates.
[DATA_EXFILTRATION]: The skill contains workflows that can expose internal API data to the public internet, which serves as a potential vector for accidental data exfiltration.
Evidence: The workflows in references/api-documentation.md and references/create-mock-servers.md utilize tools like publishDocumentation and publishMock to create publicly accessible URLs for what may be private API definitions and mock responses.
[COMMAND_EXECUTION]: The skill generates and writes executable source code to the local filesystem, which is a form of dynamic content generation that can be influenced by external data.
Evidence: The references/generate-client-code.md workflow specifically automates the creation of typed client files (e.g., .ts, .py, .go) in project directories based on the schemas and descriptions found in potentially untrusted OpenAPI specifications.

postman