The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data during its deduplication and named-entity resolution phases, which introduces a potential surface for indirect prompt injection.
Ingestion points: The skill handles data from file migrations, bulk imports, manual writes (via gbrain put or gbrain capture), and subagent outputs, as specified in the 'When This Gate Fires' section of SKILL.md.
Boundary markers: There are no specific instructions or delimiters defined to isolate external content from the agent's instructions during the 'Extract the core claim' or 'Dedup decision' steps.
Capability inventory: The skill utilizes tools for repository interaction including gbrain entity, gbrain query, gbrain search, gbrain get, gbrain link, and gbrain timeline-add. It also has the capability to delegate writes to other enrichment skills.
Sanitization: The instructions do not define any sanitization, filtering, or escaping mechanisms for the external content before it is read by the agent to assign dedup bands.

brain-ingest-gate