The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves connecting to remote challenge servers and processing binary files (ELF/PE), which provides a surface for indirect prompt injection. A malicious challenge could attempt to influence the agent's behavior via data returned by the server or embedded in binary metadata.
Ingestion points: The skill uses remote('host', port).recv() and ELF('./binary') to ingest external data.
Boundary markers: The provided exploit templates do not utilize clear delimiters or specific instructions to isolate and ignore potentially malicious content within the processed data.
Capability inventory: The agent has access to pwntools functions like process(), remote(), asm(), and ELF(), enabling it to execute code and perform network operations.
Sanitization: No sanitization or filtering is applied to the data received from remote targets before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill makes extensive use of pwntools to dynamically generate machine code (shellcode) and ROP chains. This runtime code generation is a core requirement for the skill's intended purpose of solving Pwn challenges.
[COMMAND_EXECUTION]: The skill uses the process() function from the pwntools library to execute and debug local binary files. It also describes creating and executing shell scripts (e.g., for modprobe_path exploitation in references/kernel-exploitation.md), which are standard procedures in a security research context.
[PRIVILEGE_ESCALATION]: The skill provides instructions for escalating privileges within a target environment (e.g., escaping a restricted shell in references/sandbox-escape.md or gaining root in a kernel challenge via modprobe_path or core_pattern overwrites in references/kernel.md). These techniques are intended for use against CTF targets and are consistent with the skill's primary educational purpose.

ctf-pwn