The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from external websites. * Ingestion points: Uses agent-browser snapshot and agent-browser console to read content from the target URL (SKILL.md). * Boundary markers: Absent; there are no instructions to ignore commands within the webpage content. * Capability inventory: Includes shell command execution (mkdir, cp) and extensive browser controls (click, fill, record). * Sanitization: None; the agent directly evaluates site content for issue detection.
[COMMAND_EXECUTION]: The skill uses local shell commands for environment setup and tool invocation. * It creates directories and copies report templates using mkdir and cp (SKILL.md). * It executes the agent-browser CLI tool to perform web interactions.

dogfood