The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external systems (via the RevOS MCP server) and has the ability to trigger state-changing operations, creating a surface for indirect prompt injection where malicious data in the source systems could influence the agent's behavior.
Ingestion points: The skill heavily relies on reading external data through api_read calls on resources such as action-runs, actions, tables, and connections (found in SKILL.md, references/debug.md, and references/sync-now.md).
Boundary markers: The instructions lack explicit directives for the agent to use boundary markers, delimiters, or strict validation when processing or presenting data retrieved from external sources.
Capability inventory: The skill has write access through api_write to initiate synchronization tasks on connections and queue new action-runs across potentially thousands of records (found in references/run.md and references/sync-now.md).
Sanitization: There are no specific requirements for the agent to sanitize or escape external content before interpolating it into prompts or using it as parameters for further tool calls, though it does include some human-in-the-loop confirmation steps for large-scale operations.

actions