The Agent Skills Directory

REMOTE_CODE_EXECUTION (MEDIUM): The skill includes an evaluate.js script designed to execute arbitrary JavaScript strings within the browser context via the --script flag. This allows for dynamic code execution based on agent-generated or potentially external input.
COMMAND_EXECUTION (MEDIUM): The skill's primary interface involves executing local Node.js CLI scripts. It also suggests running ./install-deps.sh and npx serve, which involve system-level command execution and potentially elevated privileges for dependency installation.
EXTERNAL_DOWNLOADS (MEDIUM): The skill requires npm install to fetch the puppeteer library and other dependencies. While a common practice, fetching unpinned dependencies from a public registry introduces a supply-chain risk if the registry or packages are compromised.
PROMPT_INJECTION (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8) because it extracts information from untrusted web sources.
Ingestion points: snapshot.js, aria-snapshot.js, and evaluate.js ingest data from arbitrary URLs.
Boundary markers: No delimiters or safety instructions are mentioned to prevent the agent from following instructions found on websites.
Capability inventory: The agent can execute local scripts, read/write session files (.browser-session.json), and perform network operations.
Sanitization: No sanitization of ingested web content is described.
PROMPT_INJECTION (LOW): The 'IMPORTANT Task Planning Notes' section contains instruction-shaping prompts that attempt to mandate specific agent behaviors (task planning and final review), which is a minor form of behavior override.

chrome-devtools