The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by documenting a pipeline that ingests untrusted data from multiple sources and processes it using an LLM for graph extraction.
Ingestion points: The remember() API accepts various inputs including strings, local file paths, s3:// URLs, and HTTP(S) URLs.
Boundary markers: The skill does not document the use of delimiters or specific instructions to the model to ignore embedded commands within the ingested data.
Capability inventory: Ingested data is processed by an LLM (extractor="llm") or GLiNER and stored in graph and vector databases.
Sanitization: No specific sanitization or filtering logic for external content is described in the ingestion process.
[EXTERNAL_DOWNLOADS]: The ingestion process automatically fetches and scrapes content from remote HTTP(S) and S3 locations. Additionally, the skill notes that the GLiNER extractor will download a pre-trained model (approximately 750 MB) from an external source upon first use.
[COMMAND_EXECUTION]: The skill relies on the presence of the git command-line utility on the system PATH for the ingestion and processing of code repositories.

cognee-ingestion