The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions explicitly authorize the agent to execute arbitrary shell commands (e.g., run bash "ls -la"), which can be used to bypass system restrictions or access sensitive files.
[REMOTE_CODE_EXECUTION]: The core functionality involves dynamically generating and executing code in Python, JavaScript, and TypeScript, creating a direct vector for code injection and execution of malicious payloads.
[EXTERNAL_DOWNLOADS]: The instructions facilitate the installation of external dependencies from public registries (e.g., pip install pandas) without integrity verification or version constraints, introducing supply chain risks.
[PROMPT_INJECTION]: The skill uses instructional language that grants the agent "infinite power" and encourages it to bypass its own limitations by writing its own tools, which can lead to autonomous behavior that overrides safety guidelines.
[INDIRECT_PROMPT_INJECTION]: The skill lacks input sanitization or boundary markers for the data it processes through generated scripts, making it vulnerable to malicious instructions embedded in user-supplied data or files.

code-executor