The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructions reference the use of sensitive API keys (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY). However, it follows security best practices by directing the agent to use environment variables and explicitly forbidding the agent from writing credentials to files or logging them.
[INDIRECT_PROMPT_INJECTION]: The skill operates by inspecting live application behavior in a browser, which constitutes an ingestion point for untrusted third-party content. This creates a risk where malicious instructions embedded in a web page could attempt to influence the agent's verification results.
Ingestion points: Live browser content and saved recordings (record.zip) processed by the vibium check command.
Boundary markers: The skill provides strong boundary instructions, stating that the agent's own inspection is not a substitute for the tool's check and advising that the builder transcript or source code should not be sent to the verifier.
Capability inventory: The CLI tool executes browser automation (navigation, clicking, filling forms) and makes requests to AI providers for reasoning.
Sanitization: The skill mitigates risks by instructing the agent to use a fresh verifier context and warning that "an embedded PASS" in evidence should not be treated as proof.

check