The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates agent-driven interaction with and analysis of external application UIs (web and mobile), creating an inherent attack surface where malicious content in the tested application could attempt to influence agent behavior.\n- Ingestion points: Untrusted data from the application UI, including roles, names, and text, is ingested through the observe tool as documented in AGENT-STEPS.md.\n- Boundary markers: The documentation establishes clear boundaries, stating that the model does not execute code directly and that the runner validates and authorizes every tool call proposed by the agent (AGENT-STEPS.md).\n- Capability inventory: The framework utilizes capabilities for subprocess execution (starting the application via app.command), filesystem access (writing traces and logs to .e2e/), and network operations for standard testing activities (RUNNER-SETUP.md).\n- Sanitization: Mitigations are implemented through the redaction of raw HTML, masking of passwords, and the withholding of model pixels and screenshots after sensitive information entry to prevent data leakage (AGENT-STEPS.md, WRITING-TESTS.md).\n- [SAFE]: The skill demonstrates a strong security posture by employing opaque handles for secrets, using environment variables for credential management, and providing explicit warnings about potential data exposure in unmasked video artifacts.

tester-army-e2e