The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install a variety of security-focused tools and libraries, including pwntools, ROPgadget, ropper, one_gadget, seccomp-tools, gdb, and pwndbg. These are installed from established services such as PyPI, RubyGems, and GitHub.
[COMMAND_EXECUTION]: The skill provides numerous shell commands for binary analysis, debugging, and environment setup. The Bash tool is used extensively to run diagnostic utilities (checksec, readelf) and to execute exploit scripts.
[INDIRECT_PROMPT_INJECTION]: The skill describes workflows for interacting with untrusted binary targets and network services, creating a potential surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through the output of target binaries (Bash tool), file reads during analysis (Read tool), and interactions with remote network services (WebFetch tool).
Boundary markers: Instructions do not consistently advise the use of delimiters or 'ignore embedded instructions' warnings when parsing output from untrusted sources.
Capability inventory: The skill uses the Bash, Read, Write, Edit, and WebFetch tools, allowing for arbitrary command execution and filesystem modifications.
Sanitization: Payloads are often constructed by parsing raw bytes from target output (e.g., leaked memory addresses) without explicit sanitization beyond ensuring valid binary formats.
[DYNAMIC_EXECUTION]: The skill includes Python templates in the scripts/ directory that dynamically generate shellcode and ROP (Return-Oriented Programming) chains based on runtime analysis of target binaries.
[PRIVILEGE_ESCALATION]: The documentation describes various techniques for achieving privilege escalation on target systems (e.g., kernel.md for kernel exploits and advanced-exploits-4.md for Windows SYSTEM escalation). These are documented as techniques to be performed on targets, not actions performed by the skill itself on the host environment.

ctf-pwn