The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to navigate to and interact with external, untrusted web pages. This exposes the agent to indirect prompt injection where malicious instructions may be hidden in page text or HTML comments.
Ingestion points: Untrusted data enters the agent context through vibium go (navigation), vibium map (DOM reading), and the output of the detection scripts in scripts/.
Boundary markers: The skill does not use explicit boundary markers to delimit untrusted page content, though it providing a detection mechanism reduces the risk of accidental obedience.
Capability inventory: The skill provides full browser interaction capabilities including clicking, typing, form filling, and capturing storage state.
Sanitization: scripts/check-injection.js provides a heuristic scanner for these attacks and includes a sanitizeSnippet function to strip terminal control characters from detected payloads before they are reported.
[COMMAND_EXECUTION]: The skill executes local binaries and Node.js scripts to perform its automation tasks.
Evidence: spawnSync is utilized in scripts/lib/vibium.js and scripts/batch.js to invoke the vibium CLI and the assert.js helper.
[EXTERNAL_DOWNLOADS]: The skill instructions describe downloading external tools and binaries necessary for its operation.
Evidence: SKILL.md instructs the user to install the vibium binary via npm and notes that the engine may download Chromium for Testing on its first execution. It also provides a workaround for ARM64 Linux that involves installing system packages like chromium-driver via apt-get.

qe-browser