The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied and third-party data from connected email marketing platforms to create tailored email sequences, introducing a risk of indirect prompt injection if subscriber fields contain untrusted content.
Ingestion points: Profile fields, names, and event details are retrieved dynamically from external platforms using tools such as list_profiles, tool_list_members, tool_list_subscribers, and tool_list_contacts in SKILL.md.
Boundary markers: The prompt instructions lack strict boundaries or delimiters to isolate subscriber data fields from the rest of the execution context, meaning malicious text inside a profile attribute could theoretically manipulate the output.
Capability inventory: The system can perform list hygiene and campaign drafting using active write capabilities like unsubscribe_profiles, tool_update_member, tool_update_subscriber, and tool_delete_contact, along with producing system findings via mcp__cogny__create_finding.
Sanitization: No data validation or verification steps are defined to clean or strip control characters and instructional strings from subscriber names or custom attributes before processing.

winback-engine