| name | ai-security |
| description | LLM and AI agent security: prompt injection, jailbreaks, agent defense, guardrails. Invoke for: "prompt injection", "LLM security", "agent security", "jailbreak defense", "AI safety audit", "system prompt leakage", "adversarial inputs", "AI pipeline security", "tool call validation", "LLM guardrails", "model security", "is my prompt safe".
|
| argument-hint | AI system, prompt, or agent pipeline to audit |
| allowed-tools | Read, Grep, Glob, WebSearch |
Skill: AI Security — LLM & Agent Defense
Category: Security
Color Team: Orange
Role
Audit LLM systems, agent pipelines, and AI workflows for security vulnerabilities unique to AI: prompt injection, jailbreaks, data exfiltration, trust boundary violations.
When to invoke
- Building or reviewing an LLM-powered application
- Agent pipeline security review
- "is my system prompt safe?"
- Tool-calling security validation
Instructions
- Review system prompts: confidential info that could leak? Injection-resistant?
- Test prompt injection: can user input override system instructions?
- Check tool call validation: does agent validate tool outputs before acting?
- Trust boundaries: does agent trust LLM output blindly? Human-in-the-loop for critical actions?
- Data exfiltration: can adversarial input cause data leakage via agent tools?
- Output validation: LLM responses sanitized before display? No SSRF via agent?
Output format
## AI Security Audit — <system> — <date>
### Prompt Injection Risk: HIGH/MEDIUM/LOW
### System Prompt Leakage: ✅/⚠️
### Tool Call Safety: ✅/⚠️
### Trust Boundaries: ✅/⚠️
### Findings & Mitigations
Example
/ai-security audit agent pipeline in src/agents/ — check prompt injection and tool trust
