The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides numerous functional payloads designed to execute arbitrary system commands within Python environments using functions such as os.system, subprocess.Popen, and eval.
[COMMAND_EXECUTION]: Instructions include specific shell commands like cat flag.txt, id, and ls to be executed via the Python interpreter to read files and identify system context.
[PRIVILEGE_ESCALATION]: The core functionality of the skill is dedicated to "escaping Python sandbox restrictions," which involves techniques to bypass security boundaries and gain broader access to the underlying operating system.
[OBFUSCATION]: The documentation describes various methods to evade keyword and character filters, such as using hex-encoded strings (\x6f\x73), chr() value concatenation, and Unicode character substitution (e.g., fullwidth ｅｖａｌ) to hide intent from automated scanners.
[DYNAMIC_EXECUTION]: The skill utilizes advanced Python internals, such as traversing the class hierarchy (().__class__.__base__.__subclasses__()), to programmatically locate and invoke restricted modules and built-in functions like os at runtime.
[EXTERNAL_DOWNLOADS]: The skill references a third-party tool hosted on GitHub (github.com/seaung/pyjailbreaker) for obtaining further automated sandbox escape payloads.
[INDIRECT_PROMPT_INJECTION]: The skill provides logic for bypassing input filters, creating a surface where an agent might be induced to execute malicious commands if it processes untrusted data. 1. Ingestion points: Python interpreter input processing via exec(input()) in reference/bypass.md. 2. Boundary markers: Absent; the skill is explicitly designed to break out of established boundaries. 3. Capability inventory: Functional implementations of os.system, subprocess.Popen, eval, and exec across all skill files. 4. Sanitization: Absent; the skill is designed to circumvent sanitization and filtering mechanisms.

pyjail