AI Security — Security & Safety

AI Security is a specialized skill for assessing AI/ML systems and LLM-based agents for security vulnerabilities such as prompt injection, jailbreak attempts, model inversion risk, data poisoning exposure, and agent tool abuse. It provides a scanner tool that detects injection signatures, scores risks, and maps findings to MITRE ATLAS techniques, ultimately producing a security assessment report with recommendations for guardrails.

engineering claude-code

📂 View skill on GitHub ↗ 🚀 How to use

✨ What it does

Scans prompts for injection signatures (direct role override, indirect injection, jailbreak persona, etc.) and assigns s
Scores model inversion risk based on access level (white-box, gray-box, black-box).
Scores data poisoning risk based on fine-tuning scope (fine-tuning, RLHF, RAG, pre-trained-only, inference-only).
Detects agent tool abuse patterns and provides mitigation strategies.
Maps findings to MITRE ATLAS techniques (e.g., AML.T0051, AML.T0020).
Provides guardrail design patterns for input validation, output filtering, and agent-specific controls.

🎯 When to use it

Before deploying an LLM in a user-facing application to check for prompt injection and jailbreak vulnerabilities.
When assessing an AI/ML system's exposure to model inversion or data poisoning risks.
When evaluating an LLM agent's tool access for potential abuse or unauthorized actions.
When needing to map AI security findings to MITRE ATLAS techniques for compliance or reporting.
When designing guardrails for an AI system to mitigate identified security risks.

🚀 How to use

Trigger the skill by asking for an AI security assessment of an LLM, classifier, or agent. Provide target type, access level, and optionally a test file of prompts. Use the provided script ai_threat_scanner.py with appropriate flags. Example prompts:

Assess the security of our LLM chatbot for prompt injection vulnerabilities.
Run an AI security scan on our classifier with white-box access.

📄 Output: A security assessment report including risk scores, findings with ATLAS mappings, and guardrail recommendations.

📦 Add this skill to Claude Code

# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills

# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering-team/skills/ai-security .claude/skills/ai-security

Skill source: engineering-team/skills/ai-security/SKILL.md

⚠️ Good to know

Requires access to test prompts or a prompt test file; for gray-box and white-box access levels, written authorization is required. The tool uses static signature matching and does not require live model access.

❓ FAQ

What is the difference between this skill and general security pen-testing?

This skill focuses specifically on AI/ML system security, such as prompt injection and model inversion, while security-pen-testing covers general application vulnerabilities like OWASP Top 10.

Does the scanner require live model access?

No, the tool uses static signature matching and assesses inputs before they reach the model, so it does not require live model access.

What does exit code 2 indicate?

Exit code 2 indicates critical findings or missing authorization for invasive access levels, meaning deployment should be blocked until issues are fixed.

🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.