The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses pwntools and shell scripts to execute binaries, interact with GDB, and run analysis tools like readelf and strings. This behavior is consistent with the skill's purpose for binary exploitation and CTF tasks.
[DYNAMIC_EXECUTION]: The toolkit includes scripts that generate executable content at runtime.
tools/patch_ld_preload.sh creates a bash wrapper script to inject shared libraries using LD_PRELOAD for exploit testing.
templates/angr_template.py uses the angr symbolic execution framework. Although static analysis tools flagged the use of eval(), the code actually calls solver.eval(), which is a safe and standard API for the angr library and not the native Python eval() function.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from binary files (e.g., through strings or decompilation). If an agent uses these tools on a malicious binary, the resulting output could potentially contain instructions aimed at the agent.
Ingestion points: Output from strings, Ghidra decompilation, and GDB registers/stack views (various files).
Boundary markers: Not explicitly defined for sanitizing binary output before agent ingestion.
Capability inventory: The skill allows for process execution, file system interaction, and network connections via pwntools templates.
Sanitization: No specialized sanitization is present for external binary data.

pwn-exploits