The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The remember command ingests data from external sources, including file paths and URLs, which are then processed by an LLM to build memory graphs. This creates an attack surface for indirect prompt injection if those sources contain adversarial instructions.
Ingestion points: cognee-cli remember accepts arbitrary text, local file paths, and URLs.
Boundary markers: No specific delimiters or warnings for embedded instructions are mentioned in the CLI context.
Capability inventory: Ingested content is stored and later retrieved/processed during recall or search operations.
Sanitization: No description of sanitization, filtering, or validation of external content is provided.
[COMMAND_EXECUTION]: The skill facilitates the execution of CLI commands for system management, including database migrations via Alembic (upgrade, downgrade) and memory state management. It explicitly notes that forget --all removes all data without a confirmation prompt, representing a safety hazard for data loss.
[CREDENTIALS_UNSAFE]: The tool manages the LLM_API_KEY environment variable and allows users to set or retrieve configuration values persisted in a local .env file. While the config get command masks secrets by default, the --show-secrets flag allows users to display raw API keys.

cognee-cli