The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides detailed instructions and ready-to-use code snippets for executing shell commands to escape restricted environments. Examples include os.system('sh') in Python, ffi.C.system('sh') in Lua, and various shell-based escapes for restricted shells (rbash) like vi, awk, and find.
[REMOTE_CODE_EXECUTION]: Includes high-risk exploit patterns for achieving arbitrary code execution through Python's pickle.loads deserialization, dynamic code object construction via types.CodeType, and Lua loadstring execution.
[PROMPT_INJECTION]: Contains methodologies for bypassing software 'jail' constraints and instruction-based sandboxes, teaching the agent how to circumvent filters such as keyword blocks, AST node restrictions, and builtin removals.
[DYNAMIC_EXECUTION]: (Internally identified as MEDIUM) The skill documents techniques for runtime code generation and execution, specifically within the PYTHON_SANDBOX_ESCAPE.md file, using compile, exec, and eval to bypass static analysis.
[OBFUSCATION]: (Internally identified as MEDIUM) Demonstrates the use of Base64 encoding (X19pbXBvcnRfXygnb3MnKS5zeXN0ZW0oJ3NoJyk=), Hex escapes (\x5f\x5f), and character concatenation (chr()) to hide sensitive keywords like __import__ and system from security scanners.

sandbox-escape-techniques