The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/lib/browser.mjs file uses createRequire and import() to dynamically resolve and load the playwright or playwright-core packages from paths relative to process.cwd(). This mechanism allows a project under audit to execute arbitrary code by providing a malicious version of these dependencies in its own node_modules directory.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially malicious content from git diff outputs and target website HTML without sufficient sanitization or the use of clear boundary markers. This attack surface could allow an adversary to embed instructions that redirect the agent's behavior during the audit process.\n
Ingestion points: Git diff output, .readyvibe/context.md files, and HTML content retrieved by the observe-runtime.mjs script.\n
Boundary markers: The instructions do not define delimiters or negative constraints to prevent the agent from following instructions found within the audited content.\n
Capability inventory: The agent is authorized to execute local Node.js scripts, interact with a headless browser, and make network requests.\n
Sanitization: The custom HTML parsing and diff observation logic do not include filters for prompt injection patterns.\n- [COMMAND_EXECUTION]: The main skill instructions direct the agent to run the node scripts/observe-runtime.mjs utility. This script automates a headless browser to perform actions such as navigation, cookie capture, and form submission on external websites, which could be abused if the audit steps are manipulated.

compliance-diff