The Agent Skills Directory

[SAFE]: No malicious patterns, such as prompt injection, obfuscation, or unauthorized data exfiltration, were identified. The skill instructions focus on standard database operations and best practices for vector storage.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for ingesting and retrieving untrusted external data for agent memory, which creates an inherent attack surface for indirect prompt injection.
Ingestion points: Untrusted content is stored in the content columns of the agent_knowledge and agent_memories tables (SKILL.md).
Boundary markers: The skill does not specify prompt delimiters within the SQL schema, but it explicitly mandates strict filtering by tenant_id and user_id at the database level to prevent data leakage between users.
Capability inventory: The skill performs database similarity searches and upsert operations.
Sanitization: The skill recommends using parameterized queries through established ORMs and tools (Prisma, Drizzle, sqlc), which effectively mitigates standard injection vulnerabilities.

pgvector