The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill is designed to enable the agent to author, store, and execute arbitrary TypeScript code using tools such as script-run and script-upsert. This allows for runtime code generation and execution which bypasses static analysis of the skill's behavior.
[COMMAND_EXECUTION]: The skill exposes the ctx.swarm.db_query functionality, which allows scripts to execute raw SQL commands against internal system tables like agent_tasks, session_logs, and agent_memory. This provides broad access to the agent's historical data and operational state.
[INDIRECT_PROMPT_INJECTION]: The skill frequently processes untrusted data from external sources, making it a surface for indirect injection attacks.
Ingestion points: Data enters the context through ctx.stdlib.fetch (external APIs), ctx.swarm.db_query (database records), and ctx.swarm.task_get (task descriptions and metadata).
Boundary markers: The documentation does not demonstrate the use of strict boundary markers or delimiters when interpolating results from these sources into further logic, though it notes that scripts run out-of-process.
Capability inventory: Scripts have access to a wide range of sensitive capabilities, including network operations (fetch), task management (task_send, task_action), and database access.
Sanitization: The platform implements a "secret scrubber" and uses Redacted wrappers for sensitive configuration values to prevent accidental exposure in logs or return values.
[DATA_EXFILTRATION]: The skill provides tools for reading sensitive internal data (tasks, memories, and logs) and includes network utilities (ctx.stdlib.fetch) that could be used to transmit this information to external endpoints.

swarm-scripts