The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill involves processing inbound MMS payloads (MESSAGE_INBOUND) which contain content generated by end-users, such as text and media URLs. This creates a surface for indirect prompt injection if not handled carefully.
Ingestion points: Untrusted data enters the context through inbound message payloads and webhook triggers described in the 'Common Patterns' and 'Security' sections of SKILL.md.
Boundary markers: The skill explicitly instructs the agent to treat this data as untrusted and provides warnings to avoid interpolating it into prompts or code.
Capability inventory: The skill utilizes network capabilities via curl and API requests to Sinch endpoints, as well as potential interaction with external media URLs.
Sanitization: The skill recommends verifying webhook signatures (HMAC validation) and sanitizing content before processing, referencing the sinch-conversation-api skill for implementation details.

sinch-mms