The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to manage and merge chat message history from external HTTP requests into a persistent store, creating a vector for indirect prompt injection if malicious data is saved to the transcript.
Ingestion points: The chatParamsFromRequest function in the POST handler and the reconstructChat function in the GET handler ingest untrusted data from request bodies and query parameters.
Boundary markers: No explicit delimiter or instruction filtering is demonstrated in the message merging logic; it relies on standard model message arrays.
Capability inventory: The skill utilizes a persistence adapter to perform write operations to storage backends (messages, runs, interrupts).
Sanitization: The implementation examples do not include explicit validation or sanitization of the message content before storage.
[DATA_EXFILTRATION]: The documentation explicitly identifies a potential data exposure vulnerability where chat transcripts could be retrieved by guessing a threadId. It provides a defensive coding pattern using an authorize callback to verify user ownership before serving history, which mitigates the risk of unauthorized data access.

ai-persistence/server