The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/recon.py utilizes subprocess.run to invoke the standard c++filt utility for demangling C++ symbols. This implementation is secure as it avoids shell=True, uses a list of arguments, and passes the symbol name via standard input with a 5-second timeout, preventing shell injection or resource exhaustion.
[INDIRECT_PROMPT_INJECTION]: The skill inherently processes untrusted data (binary files and disassembly). However, it includes robust defensive engineering by providing explicit 'Non-Negotiable' rules in SKILL.md and a dedicated reference file references/untrusted-binary-content.md. These instructions warn the agent that binary-derived content (strings, decompiler comments, log lines) is attacker-controlled data that must never be interpreted as instructions or used to authorize tool calls.
[EXTERNAL_DOWNLOADS]: The skill recommends several well-known third-party libraries for binary analysis, including lief, capstone, pwntools, and triton-library. These are standard, reputable tools in the security research community. The documentation provides clear guidance on resolving a name collision with OpenAI's unrelated triton package by using a dedicated virtual environment.

x86-cpp-reversing