The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of automation pipelines ('Factory lines') and jobs that ingest untrusted data from external sources such as GitHub pull requests, Slack messages, and HTTP webhooks. This architecture introduces a risk of indirect prompt injection, where an external attacker could influence agent behavior by submitting malicious content to these integrated services.
Ingestion points: Data enters the agent context through integration triggers (GitHub, Slack, Linear) and incoming webhooks defined in factory.md and webhooks.md.
Boundary markers: The documentation does not explicitly mandate boundary markers or sanitization for raw input data within the manifest examples, although it encourages the use of run_agent steps for processing.
Capability inventory: The platform supports executing arbitrary shell commands (exec steps), file system writes, and network operations through the gateway as documented in jobs.md and sandbox-lifecycle.md.
Sanitization: No explicit sanitization or filtering logic is prescribed for handling external input before it is passed to agent prompts.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill references the official islo-labs/islo-agents GitHub repository and encourages installing the vendor's own SDKs (@islo-labs/sdk and islo on PyPI). These references target the platform's official resources and are documented neutrally as intended functionality.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to use the islo use and islo job run commands to execute tasks within managed cloud sandboxes. This is the primary function of the platform and is performed within an isolated environment.

using-islo