The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill navigates to external, potentially untrusted URLs and inventories DOM elements, network responses, and console logs. This represents a significant ingestion surface for malicious instructions hidden in web pages (e.g., hidden text, meta tags, or mock data payloads) that could attempt to influence the agent's behavior during the audit. The skill attempts to mitigate this with a multi-phase workflow and explicit human-in-the-loop checkpoints for interactions.
[COMMAND_EXECUTION]: The skill uses Playwright MCP (via browser_navigate) and supports executing shell commands for database verification (e.g., psql, mysql, supabase). While these are intended for read-only audits, they represent powerful capabilities that could be abused if the agent is misled by injected instructions.
[SAFE_PRACTICE]: The skill explicitly instructs the agent to use placeholder credentials and dedicated test accounts, avoiding the use of real user secrets for form testing. It also includes a heuristic for skipping destructive-looking UI elements.

mock-hunter