sandbox-escape-techniques

Fail

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: HIGHOBFUSCATIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
  • [OBFUSCATION]: The file PYTHON_SANDBOX_ESCAPE.md contains a Base64-encoded payload X19pbXBvcnRfXygnb3MnKS5zeXN0ZW0oJ3NoJyk= which decodes to the executable command __import__('os').system('sh'). This is demonstrated within an eval(base64.b64decode(...)) block, providing a functional template for hiding malicious code execution.
  • [COMMAND_EXECUTION]: The skill documents and provides code for numerous shell execution methods intended to bypass environment restrictions, including os.system('sh'), ffi.C.system("sh"), and execl("/bin/sh", "sh", NULL). It also describes using terminal ioctl (TIOCSTI) to inject keystrokes into a parent shell.
  • [PRIVILEGE_ESCALATION]: The skill provides detailed instructions for escaping various forms of isolation, including chroot jails via double-chrooting, Docker/container escape via nsenter and unshare to enter host namespaces, and kernel-level bypasses using ptrace for syscall injection.
  • [DYNAMIC_EXECUTION]: The skill promotes the use of dynamic execution primitives such as eval(), exec(), pickle.loads(), and types.CodeType specifically to bypass static analysis filters and AST-based security scanners.
  • [PROMPT_INJECTION]: The AI LOAD INSTRUCTION sections in SKILL.md, PYTHON_SANDBOX_ESCAPE.md, and SECCOMP_BYPASS.md instruct the agent to adopt a persona capable of "breaking out" of sandboxes and achieving "unrestricted code execution," which encourages the agent to prioritize these techniques over default safety constraints.
  • [INDIRECT_PROMPT_INJECTION]: The skill creates a high-risk capability surface that could be exploited if the agent processes untrusted data using these techniques.
  • Ingestion points: Files SKILL.md, PYTHON_SANDBOX_ESCAPE.md, and SECCOMP_BYPASS.md are loaded into the agent's context as authoritative instructions.
  • Boundary markers: Absent; there are no instructions to the agent to treat these techniques as theoretical or to maintain safety boundaries while applying them.
  • Capability inventory: The skill provides instructions for exec, eval, ptrace, nsenter, unshare, and raw filesystem access via /proc/1/root.
  • Sanitization: No sanitization or validation logic is provided for the shell commands or payloads generated by the skill.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 9, 2026, 11:59 AM