The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to supervise browser interactions, which involves ingesting untrusted data from external websites (e.g., page content, forms, and blockers). While the skill acts as a safety layer, the ingestion of external data presents a surface for indirect prompt injection.
Ingestion points: Processes target context and user requests related to browser tasks, including capturing page traces and DOM notes from potentially attacker-controlled websites.
Boundary markers: Explicitly defines 'auth boundaries', 'prepared-vs-observed boundaries', and 'not-evidence boundaries' to distinguish between prepared guidance and actual execution results.
Capability inventory: Includes capabilities for browser automation (click, login, fill form) and runtime telemetry recording via the omh CLI.
Sanitization: Implements 'confirmation gates' for destructive actions (payment, destruction) and requires 'exact approval' for any browser effects via docs/BROWSER-EFFECTS.md.
[COMMAND_EXECUTION]: The skill documentation includes examples of shell commands used for runtime telemetry.
Evidence: The skill references omh runtime record --skill browser-operator --harness browser-operator --status started for logging delegation results.

omh-browser