The Agent Skills Directory

[COMMAND_EXECUTION]: Provides detailed payloads and one-liners for executing shell commands through various languages and tools, including Python's os.system and subprocess.Popen, Lua's os.execute, and restricted shell escapes using vi, awk, and find.
[OBFUSCATION]: Documents methodologies for bypassing keyword-based security filters. Techniques include Base64 encoding (e.g., base64.b64decode), hex and Unicode escape sequences in strings, and string construction using chr() to hide function names like __import__ and system.
[REMOTE_CODE_EXECUTION]: Details high-impact exploitation vectors such as Mojo IPC bugs for browser sandbox escapes, io_uring for bypassing Seccomp filters, and insecure Python pickle deserialization to achieve code execution.
[PRIVILEGE_ESCALATION]: Outlines procedures for escaping restricted environments, including double chroot escapes, Linux namespace escalation (unshare), and Seccomp architecture confusion (using 32-bit int 0x80 syscalls on 64-bit systems).
[DYNAMIC_EXECUTION]: Demonstrates the use of eval() and exec() for runtime code evaluation and the manual construction of Python code objects (types.CodeType) to bypass static analysis tools and Abstract Syntax Tree (AST) restrictions.
[INDIRECT_PROMPT_INJECTION]: The skill provides the agent with a significant repository of exploitation primitives which could be leveraged if the agent is instructed to process untrusted external data.
Ingestion points: The skill instructions and methodologies are loaded into the agent context via SKILL.md, PYTHON_SANDBOX_ESCAPE.md, and SECCOMP_BYPASS.md.
Boundary markers: The skill uses AI LOAD INSTRUCTION markers to define context, but does not provide explicit sanitization or boundary markers for runtime data processing.
Capability inventory: The skill grants the agent knowledge of file system access, process injection (ptrace), command execution, and network exfiltration techniques.
Sanitization: No active sanitization of external content is implemented within the playbook's documentation-style files.

sandbox-escape-techniques