The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted customer data, creating a surface for indirect prompt injection. 1. Ingestion points: The agent reads customer database exports, purchase logs, and webhook payloads (user_signup, checkout_started) as described in SKILL.md and references/contacts.md. 2. Boundary markers: The instructions lack explicit delimiting or escaping requirements for untrusted data fields, which could allow the agent to follow instructions embedded in customer traits. 3. Capability inventory: The skill is configured to use bash, run_command, and file modification tools (write_to_file, replace_file_content) for deliverability auditing and configuration management. 4. Sanitization: The skill emphasizes PII protection and GDPR/TCPA compliance, but does not specify procedures for sanitizing natural language instructions that might be contained within custom customer properties or metadata.

crm