The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web pages via the inspect-ui.mjs and scan-ui.mjs scripts. An attacker could embed malicious instructions in the HTML/CSS of a site being scanned to influence agent behavior.
Ingestion points: scripts/probe-ui.mjs (via probeUI and checkpoint functions) and scripts/inspect-element.mjs (via elementFacts) extract data from target URLs.
Boundary markers: The scripts do not use explicit delimiters or 'ignore' warnings for extracted content.
Capability inventory: The agent can execute shell commands (bun), write local files, and perform browser automation with network access.
Sanitization: elementFacts filters out text nodes within input, textarea, and select elements to reduce the risk of PII leakage.
[COMMAND_EXECUTION]: The skill requires the execution of multiple JavaScript scripts using the bun runtime. These scripts perform browser automation and interact with the local filesystem to save results.
[DYNAMIC_EXECUTION]: The inspection toolkit utilizes dynamic loading and evaluation.
Dynamic Loading: scripts/inspect-ui.mjs loads @playwright/test from a path derived from process.cwd(), which is a dynamic path.
Runtime Evaluation: Standard browser automation patterns using page.evaluate() are employed to execute extraction logic within the browser context.

ui-improvement