← ClaudeAtlas

ctf-ai-securitylisted

AI and LLM security CTF recipes. Load for any AI hacking category challenge, prompt injection, jailbreak, MCP tool poisoning, agent tool abuse, provided model files, or AI powered filter bypass.
bilal-spotted/claude-code-ctf-rig · ★ 3 · AI & Automation · score 71
Install: claude install-skill bilal-spotted/claude-code-ctf-rig
# CTF AI Security Recipes All tools below live in an isolated virtual environment at /home/YOUR_USERNAME/ai_tools_venv, kept separate so version requirements here never fight with pwntools or the rest of the core toolkit. Always invoke by full path, /home/YOUR_USERNAME/ai_tools_venv/bin/TOOLNAME, never by bare name. ## Frameworks - OWASP LLM Top 10, OWASP MCP Top 10 - MITRE ATLAS (atlas.mitre.org), the AI equivalent of ATT&CK ## Automated probing tools Garak, NVIDIA's scanner (v0.9.0.4 here). Console launcher confirmed present at the full path: /home/YOUR_USERNAME/ai_tools_venv/bin/garak --model_type TARGET_TYPE --model_name NAME --probes leakreplay (system prompt / training data leak) /home/YOUR_USERNAME/ai_tools_venv/bin/garak --model_type TARGET_TYPE --model_name NAME --probes dan,encoding (jailbreak and filter bypass) The launcher is a thin shim over the module, so /home/YOUR_USERNAME/ai_tools_venv/bin/python3 -m garak <same args> is an exact equivalent if the launcher is ever missing after a venv rebuild. PyRIT (v1.1.0 here), best for Crescendo escalation, used as a Python library from the venv's own python, not a standalone CLI. In 1.x the old orchestrators are "attacks" under pyrit.executor.attack: /home/YOUR_USERNAME/ai_tools_venv/bin/python3 -c "from pyrit.executor.attack import PromptSendingAttack, CrescendoAttack" (PromptSendingAttack replaces the pre-1.0 PromptSendingOrchestrator; CrescendoAttack drives multi-turn escalation.) Promptfoo lives outside the