Giving LLMs a crash course in SQLi, ROP chains, and heap feng shui

What it does

ctf-skills is a curated knowledge base packaged for the Agent Skills spec, designed to turn AI coding assistants like Claude Code or Gemini CLI into CTF teammates. It catalogs specific attack techniques—from PHP filter-chain LFI to House of Apple heap exploitation—as structured context that agents can invoke during challenges. The repo also ships a bash installer (install_ctf_tools.sh) that preps environments with category-specific tooling and supports dry-run verification.

The interesting bit

This isn’t generic “think like a hacker” prompting; it is an encyclopedia of esoteric vulnerability recipes. The web skill enumerates distinct bypasses and chains including “MySQL REGEXP byte-by-byte oracle with backtick comment bypass” and “PHP7 OPcache binary webshell + LD_PRELOAD disable_functions bypass.” It effectively compresses years of CTF tribal knowledge into agent-loadable chunks.

Key highlights

• Covers the full CTF spectrum: web exploitation, binary pwn, AI/ML attacks, crypto, reverse engineering, forensics, and OSINT per the repository description
• Extreme granularity in the visible skills: the web module lists specific techniques spanning SQLi, XSS, SSTI, SSRF, JWT injection, prototype pollution, and GraphQL CSRF, while the pwn module covers ROP, SROP, heap exploitation, and exotic gadget chains
• Bundles a bash setup script (install_ctf_tools.sh) with dry-run, verify, and category-filtered modes for pre-competition tooling
• Compatible with any Agent Skills runner, including Claude Code, Codex CLI, and Friday Studio
• 2,659 stars and actively tagged across major AI agent ecosystems

Caveats

• The published README is truncated, so the specific files and descriptions for crypto, reverse engineering, forensics, and OSINT skills aren’t visible in the provided source
• Tool installation relies on bash, apt, brew, and other Unix-centric package managers; Windows support is unclear
• This is knowledge scaffolding for AI agents, not a standalone exploitation framework or traditional Python library

Verdict

CTF competitors already using AI assistants like Claude Code get a dense context pack of exploit recipes; if you don’t run agent-based workflows or just want standalone tools, this is essentially a very well-organized set of prompts.