The Agent Skills Directory

[OBFUSCATION]: The skill documents and utilizes several techniques to hide malicious commands and keywords. Specifically, PYTHON_SANDBOX_ESCAPE.md contains an execution chain using Base64: eval(base64.b64decode('X19pbXBvcnRfXygnb3MnKS5zeXN0ZW0oJ3NoJyk=')). It also provides instructions for constructing dangerous strings using chr() values, hex escapes, and Unicode escapes to bypass static keyword filters.
[COMMAND_EXECUTION]: The skill provides numerous methods to execute arbitrary shell commands on the underlying system. Examples include os.system('sh'), subprocess.Popen(['sh']), and using Lua's FFI to call system("sh"). Additionally, it includes a comprehensive list of restricted shell (rbash) escapes using common utilities like vi, awk, and find.
[DYNAMIC_EXECUTION]: The skill relies on and teaches the use of dynamic code execution functions to bypass security controls. This includes extensive use of eval() and exec() in Python, loadstring() in Lua, and the manipulation of Python CodeType objects to execute arbitrary bytecode directly.
[PRIVILEGE_ESCALATION]: The skill documents methods to transition from restricted environments to host access. This includes 'Double Chroot' escape techniques, namespace escape techniques (e.g., unshare, nsenter) to access host resources from a container, and browser sandbox escape vectors targeting Mojo IPC.
[DATA_EXFILTRATION]: The skill provides instructions for accessing sensitive local files through non-standard interfaces, such as reading /etc/passwd using pathlib, codecs, or linecache. It also includes payload examples for exfiltrating sensitive data to an external server (http://attacker/) via curl and sendfile.
[PROMPT_INJECTION]: The skill uses 'AI LOAD INSTRUCTION' and an 'Expert Attack Playbook' persona to frame dangerous activities as expected behavior, encouraging the agent to act as an attacker capable of bypassing security filters and restrictions.

sandbox-escape-techniques