The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the implementation of hooks that process untrusted data from external HTTP requests and responses, creating a potential surface for indirect prompt injection attacks.
Ingestion points: Implementation of BeforeRequestHook, AfterSuccessHook, and AfterErrorHook in the src/hooks/ directory.
Boundary markers: There are no instructions or examples provided for using delimiters or markers to isolate external data from agent processing.
Capability inventory: The skill allows for arbitrary code execution within hooks, including header manipulation and network calls using fetch.
Sanitization: No sanitization or validation logic is suggested for handling the content of incoming responses or outgoing requests.
[DATA_EXFILTRATION]: The skill provides example code that performs network operations to an external domain, which could be used to send data out of the environment.
Evidence: An implementation example in SKILL.md demonstrates using the fetch API to send event data to https://telemetry.example.com/events.
[COMMAND_EXECUTION]: The skill includes instructions for running the Speakeasy CLI tool to regenerate the SDK.
Evidence: SKILL.md contains the command speakeasy generate sdk -s openapi.yaml -o . -l typescript for generating code based on an OpenAPI specification.

customize-sdk-hooks