The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection. It directs the agent to ingest and reason about untrusted content from external websites using browser tools.\n
Ingestion points: Data enters the agent's context through browser_snapshot (DOM structure), browser_vision (visual rendering), and browser_console (execution logs) as defined in the systematic workflow in SKILL.md.\n
Boundary markers: The instructions do not include markers or directives to distinguish between the skill's logic and potentially malicious instructions that might be embedded in the target website's content.\n
Capability inventory: The agent has capabilities to write files (reports and screenshots) and perform interactive browser actions like clicking and typing.\n
Sanitization: There is no requirement for the agent to sanitize or validate the content retrieved from the target URL before processing it.\n- [COMMAND_EXECUTION]: The workflow includes instructions for the agent to execute shell commands, specifically for setting up network tunnels using cloudflared and redirecting logs to the /tmp directory (e.g., cloudflared tunnel and > /tmp/quicktunnel.log).\n- [EXTERNAL_DOWNLOADS]: The skill mentions the use of cloudflared, an external tool from Cloudflare (a well-known service), to facilitate the testing of local applications. This involves running an external binary to create a public gateway for local development traffic.

dogfood