The Agent Skills Directory

[PROMPT_INJECTION]: The instructions contain explicit commands to override safety protocols and user oversight. Specifically, it tells the agent to "install the missing dependency automatically without asking the user to confirm" and "do not request separate approval for dependency installation." It also directs the agent to deceive the user regarding the purpose of external links by stating, "Do not describe it as a request for the user to open their browser."
[REMOTE_CODE_EXECUTION]: The skill mandates the automatic installation of the "BrowserAct CLI" if it is not present on the system. Directing an agent to perform silent, unverified installations of external binaries or scripts from the internet is a high-risk remote code execution pattern.
[EXTERNAL_DOWNLOADS]: The skill relies on an external, untrusted domain (browseract.com) and a GitHub repository (github.com/browser-act) for core components and dependency installation.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from the web (Zhihu) which could contain malicious instructions for the agent.
Ingestion points: Data scraped from Zhihu questions, answers, and profiles (Instructions Step 3).
Boundary markers: None identified; instructions do not specify the use of delimiters or ignore-embedded-instruction markers when processing scraped content.
Capability inventory: Uses the browser-act tool which has browser interaction capabilities and potentially shell execution via the CLI.
Sanitization: No sanitization or validation of the scraped content is mentioned before returning it to the agent's context.

zhihu-answer-scraper