Skip to main content
skill CX Codex

ai-threat-testing — Claude Code skill: OWASP LLM vulnerability testing

ai-threat-testing is a Claude Code skill that tests LLM applications for OWASP Top 10 vulnerabilities with 10 specialized agents.

MIT 210 installs ★ 632 updated 2026-07-29

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add transilienceai/communitytools --skill ai-threat-testing -a claude-code
Open on GitHub

What ai-threat-testing does

The skill deploys agents targeting each OWASP LLM Top 10 category (2025 edition): prompt injection and filter evasion, training data and RAG leakage, dependency CVEs and model provenance, data poisoning, code/XSS injection downstream, tool exploitation, system prompt extraction, embedding attacks, hallucination, and token flooding. It includes playbooks for each technique, a mapping catalogue, evidence collection, and CVSS-scored reporting.

Integrates with /pentest for unified security assessment. Three modes: full audit of all ten categories (4–8 hours), focused testing on one category (1–3 hours), or supply chain audit (2–4 hours). Captures screenshots, network logs, API responses, and generates proof-of-concept scripts.

Requires written authorization and preferably an isolated test environment; all findings are documented with reproducible PoCs.

What it helps with

  • find prompt injection and instruction override vulnerabilities
  • detect training data leakage and cross-tenant RAG extraction
  • audit dependencies for CVEs and verify model integrity
  • identify unsafe output handling and downstream code injection
  • discover tool abuse and unauthorized agent actions

Example prompts for the agent

  • Run a full OWASP LLM Top 10 assessment on the LLM application
  • Test only for prompt injection and system prompt extraction
  • Audit the supply chain: CVEs, plugins, and model provenance

How to install ai-threat-testing

1Skills CLI (Vercel)

npx skills add transilienceai/communitytools --skill ai-threat-testing -a claude-code

2Manually: put the files in place

~/.claude/skills/ai-threat-testing/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/ai-threat-testing/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about ai-threat-testing

How do I install ai-threat-testing in Claude Code?

Run npx skills add transilienceai/communitytools --skill ai-threat-testing -a claude-code, then open ~/.claude/skills/ai-threat-testing/SKILL.md.

What does ai-threat-testing do?

Tests LLM applications against OWASP LLM Top 10 (2025 edition) vulnerabilities: injection, data disclosure, poisoning, supply chain attacks, and more. Ten agents execute tests, collect evidence, and produce a report with CVSS scores and reproducible proofs of concept.

Which platforms does ai-threat-testing support?

The skill works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, and OpenCode. Each platform has its own installation command.

Source text

SKILL.md · MIT
---
name: ai-threat-testing
description: Offensive AI security testing and exploitation framework. Systematically tests LLM applications for OWASP Top 10 vulnerabilities including prompt injection, model extraction, data poisoning, and supply chain attacks. Integrates with pentest workflows to discover and exploit AI-specific threats.
---

# AI Threat Testing

Test LLM applications for OWASP LLM Top 10 vulnerabilities using 10 specialized agents. Use for authorized AI security assessments.

## Quick Start

```
1. Specify target (LLM app URL, API endpoint, or local model)
2. Select scope: Full OWASP Top 10 | Specific vulnerability | Supply chain
3. Agents deploy, test, capture evidence
4. Professional report with PoCs generated
```

## Coverage — OWASP LLM Top 10, 2025 edition

**Which file addresses which category is decided by
[`reference/catalog/llm-top10-2025.json`](reference/catalog/llm-top10-2025.json), not by the
filename.** The `llmNN-` prefixes on disk predate the 2025 renumbering and no longer match; the
content is correct, the labels were not. Cite an id only with its edition (`LLM06:2025`), because a
bare `LLM06` means two different categories depending on which edition the reader assumes.

| Category | Attack surface |
|---|---|
| `LLM01:2025` Prompt Injection | Direct and indirect injection, instruction override, filter evasion |
| `LLM02:2025` Sensitive Information Disclosure | Training-data and cross-tenant RAG leakage, canary verification |
| `LLM03:2025` Supply Chain | Dependency CVEs, model provenance, malicious serialized models |
| `LLM04:2025` Data and Model Poisoning | Backdoor triggers, membership inference, behavioural anomalies |
| `LLM05:2025` Improper Output Handling | Code/XSS injection downstream, unsafe deserialization |
| `LLM06:2025` Excessive Agency | Tool/plugin abuse, privilege escalation, unauthorised actions — the category that matters for **agents** rather than chatbots |
| `LLM07:2025` System Prompt Leakage | **Gap — no playbook yet.** See the catalogue: what the prompt *contains* is a separate finding from whether it can be extracted |
| `LLM08:2025` Vector and Embedding Weaknesses | RAG injection, retrieval manipulation, embedding inversion |
| `LLM09:2025` Misinformation | Hallucination and confidence manipulation where output is relied upon |
| `LLM10:2025` Unbounded Consumption | Token flooding, cost impact, and **model extraction/theft** (2025 treats extraction-by-query as a consumption problem) |

Two classes are testable but are **not** OWASP categories, so they carry local `TX-` ids rather than
an invented `LLMnn`: monitoring evasion / forensic gaps, and adversarial perturbation of non-text
input. `tools/test_llm_numbering.py` enforces that separation.

## Workflows

**Full Assessment** (4-8 hours):
```
- [ ] Reconnaissance
- [ ] Deploy all 10 agents
- [ ] Execute exploits
- [ ] Capture evidence
- [ ] Generate report
```

**Focused Testing** (1-3 hours):
```
- [ ] Select a category from the catalogue (LLM01:2025 .. LLM10:2025, or a TX- local class)
- [ ] Deploy agent
- [ ] Execute techniques
- [ ] Document findings
```

**Supply Chain Audit** (2-4 hours):
```
- [ ] Inventory dependencies
- [ ] Scan CVEs
- [ ] Test plugins/APIs
- [ ] Verify model provenance
```

## Integration

Enhances `/pentest` with AI-specific testing:
- Traditional pentesting + AI threat testing = complete security assessment
- Chain vulnerabilities across traditional and AI vectors
- Unified reporting with CVSS scores

## Key Techniques

**Prompt Injection**: Instruction override, system prompt extraction, filter evasion
**Model Extraction**: Query sampling, token analysis, membership inference
**Data Poisoning**: Behavioral anomalies, backdoor triggers, bias analysis
**DoS**: Token flooding, recursive expansion, context exhaustion
**Supply Chain**: CVE scanning, plugin audit, model verification
**MCP Tool Abuse**: MCP server inspectors/debuggers often expose `/api/mcp/connect` or similar endpoints that accept `serverConfig` with arbitrary `command` parameters — unauthenticated RCE. Check for MCP Inspector, MCP Playground, or any MCP debugging UI on non-standard ports (6274, 3000, etc.).

## Evidence Capture

All agents collect: screenshots, network logs, API responses, errors, console output, execution metrics.

## Reporting

Automated reports include: executive summary, detailed findings (CVSS scores), PoC scripts, evidence, remediation guidance.

## Critical Rules

- Written authorization REQUIRED before testing
- Never exceed defined scope
- Test in isolated environments when possible
- Document all findings with reproducible PoCs
- Follow responsible disclosure practices

## Integration

- Integrates with `/pentest` skill for comprehensive security testing
- AI-specific vulnerability knowledge in `/AGENTS.md`
- Attack playbooks in `reference/llm0X-*.md`

© transilienceai · MIT · transilienceai/communitytools

Similar skills

All skills for Security →

From the same set42

The whole set →
hackthebox A skill for HackTheBox competitions that spins up machines, handles credentials, delegates attacks to subagents, and posts breach reports to Slack. CX Codex +1 213installs server-side server-side guides security testing for SSRF, HTTP request smuggling, path traversal, and deserialization flaws in server code. CX Codex +1 213installs authentication This skill checks login flows, tokens, 2FA, CAPTCHA, and bot detection. Use it to examine auth bypasses and weaknesses in authentication controls. CX Codex +1 212installs infrastructure The infrastructure skill tests network services, protocols, and perimeter security. Use it for authorized assessments that need structured checks and captured evidence. CX Codex +1 211installs injection A skill for Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, and OpenCode that tests SQL, NoSQL, OS Command, SSTI, XXE, and LDAP/XPath injection vulnerabilities. Runs from input discovery through exploitation to PoC capture. CX Codex +1 210installs client-side The client-side skill tests web apps and SPAs for XSS, CSRF, CORS errors, clickjacking, DOM attacks, and Prototype Pollution. CX Codex +1 209installs reverse-engineering reverse-engineering guides disassembly and decompilation of ELF, PE, and Mach-O files to recover logic, emulate custom bytecode, and defeat binary obfuscation. CX Codex +1 203installs dfir dfir — a skill for analyzing Windows event logs, network captures, and filesystem artifacts during security incident investigations. Reconstructs Active Directory attack chains and builds timelines across multiple evidence sources. CX Codex +1 202installs

Project sponsors

Companies that keep this analytics open