The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data that influences the instructions of generated agents.
Ingestion points: The skill parses external strings π.tasks and π.roles provided at runtime to define the swarm's behavior.
Boundary markers: There are no explicit delimiters or warnings to the platform to ignore potential malicious instructions within the role.instructions field provided in the roles input.
Capability inventory: The skill possesses significant capabilities, including the creation of persistent agents (agents.create), the ability to send messages to those agents (agents.tell), and the power to perform atomic writes and broadcasts to a shared state mesh (mesh.put, mesh.publish).
Sanitization: While the input is parsed as JSON, the logic does not perform validation or sanitization of the natural language instructions assigned to the created actors.

fabric-swarm