The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends a 'Quick Install' method that pipes remote scripts directly into system interpreters from an untrusted domain (fudankw.cn).
Linux/macOS: bash -c "$(curl -fsSL http://fudankw.cn:9000/files/ga_install.sh)"
Windows: powershell -ExecutionPolicy Bypass -c "... irm http://fudankw.cn:9000/files/ga_install.ps1 | iex"
[COMMAND_EXECUTION]: The framework is explicitly designed to grant the agent 'system-level control over computers,' including the ability to execute arbitrary shell commands via a 'Terminal Execution' tool.
[PRIVILEGE_ESCALATION]: The Windows installation command uses -ExecutionPolicy Bypass, which explicitly overrides system security settings to allow the execution of unsigned scripts.
[PERSISTENCE]: The skill provides usage examples that instruct the agent to create recurring tasks and persistence mechanisms, such as: 'Set up a cron job that runs every day at 9 AM to backup ~/Documents'.
[DYNAMIC_EXECUTION]: The skill implements 'automatic skill crystallization,' which involves the agent generating new Python code at runtime, saving it to a local library, and subsequently executing it. It also provides a 'Python REPL' tool for interactive code execution.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a wide attack surface for indirect injection by processing untrusted external data.
Ingestion points: The agent is designed to read Gmail messages, WeChat chats, Alipay transaction history, and browse arbitrary websites (Amazon, TechCrunch).
Boundary markers: There are no markers or instructions provided to the agent to ignore instructions embedded within this external data.
Capability inventory: The agent has access to terminal, filesystem, network, mouse/keyboard control, and adb (Android Debug Bridge).
Sanitization: No sanitization or validation of external content is described before the data is processed by the LLM.

genericagent-self-evolving-ai-agent