The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download a Python script (gef.py) from a third-party GitHub repository (hugsy/gef) directly into the GDB initialization file (~/.gdbinit). This script is executed automatically by GDB with the user's privileges.
[REMOTE_CODE_EXECUTION]: The instructions recommend cloning the pwndbg repository from GitHub and executing its setup script (./setup.sh), which performs arbitrary system modifications and dependency installations.
[PRIVILEGE_ESCALATION]: The skill explicitly provides a command to modify the kernel's ptrace_scope using sudo sysctl -w kernel.yama.ptrace_scope=0. This reduces the system's security posture by allowing processes to attach to and debug other processes.
[PERSISTENCE]: By modifying the ~/.gdbinit file, the skill ensures that third-party code remains persistent and is executed every time the GDB debugger is invoked in the environment.
[COMMAND_EXECUTION]: The skill makes heavy use of high-privilege shell commands, including package managers (apt, dnf, pacman) and direct kernel parameter modification.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and display data from untrusted binaries (e.g., via x/s, p $rax, and find). This data is ingested into the agent's context without sanitization or boundary markers.
Ingestion points: GDB output from memory reads, register inspection, and string searches (SKILL.md).
Boundary markers: Absent.
Capability inventory: Shell execution through GDB, file system writes via memory manipulation, and network access.
Sanitization: Absent.
[EXTERNAL_DOWNLOADS]: The skill relies on several external sources for its core functionality, including GitHub and the Python Package Index (PyPI).

re-gdb