The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides an installation command that downloads a shell script from an untrusted GitHub repository and pipes it directly to the shell for execution.
Evidence: curl -fsSL https://raw.githubusercontent.com/nashsu/opencli-rs/main/scripts/install.sh | sh in SKILL.md.
[PERSISTENCE]: The skill includes instructions to modify the user's shell configuration file, which allows for persistent execution of code across new terminal sessions.
Evidence: opencli-rs completion bash >> ~/.bashrc in SKILL.md.
[COMMAND_EXECUTION]: The skill can control local desktop applications and execute commands through system-level CLI tools.
Evidence: Control of Electron applications such as Cursor, ChatGPT, Notion, and Discord.
Evidence: Passthrough execution for gh, docker, and kubectl CLI tools.
Evidence: Ability to register and execute arbitrary third-party CLI tools via opencli-rs register.
[EXTERNAL_DOWNLOADS]: The skill downloads binaries and source code from an untrusted external repository (github.com/nashsu).
Evidence: Invoke-WebRequest for Windows binaries and git clone for source code downloads.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from over 55 websites, creating a surface for injection attacks where malicious web content could influence the agent's behavior.
Ingestion points: Data fetching from sites like Twitter, Reddit, YouTube, and Bilibili.
Boundary markers: Absent.
Capability inventory: Remote code execution, shell persistence, application control, and network access.
Sanitization: Absent.
[DATA_EXFILTRATION]: The skill accesses sensitive user data by reusing browser sessions via a Chrome extension and reading content from desktop applications like Notion and Cursor.
[DYNAMIC_EXECUTION]: The skill generates and executes custom scraping pipelines (adapters) at runtime based on the structure of external websites.
Evidence: Use of opencli-rs generate to create YAML adapters from web targets.

opencli-rs-web-scraper