OmniRed

Multi-AI Offensive Security Skills Library

The first offensive security skills library built for all four major AI platforms: Claude, ChatGPT, Gemini, and Microsoft Copilot.

Author: Sunil Gentyala, Independent Researcher License: Apache-2.0

---

Why OmniRed

Every existing offensive security skills library targets a single AI. Security operators use multiple AI platforms depending on context, client environment, and task type. OmniRed gives you the same expert methodology regardless of which AI you are using.

Beyond platform breadth, OmniRed introduces three categories that no other library covers:

Category	What it covers	Why it matters
ai-native/	Prompt injection, jailbreaking, model extraction, system prompt leak	Direct attacks against AI systems as targets
mcp/	Tool poisoning, rug pull, context injection, server impersonation	Attacks on Model Context Protocol agent pipelines
llm-pipeline/	RAG poisoning, embedding attacks, retrieval manipulation	Attacks on the data layer feeding LLMs

These categories were developed alongside published research in MCP security (ContextGuard, ICCBI 2026) and the ARGUS agentic red-team scanner.

---

Platforms

Platform	Format	Install
Claude (Claude Code)	SKILL.md plugin	See below
ChatGPT	Custom GPT instructions	chatgpt/PLATFORM.md
Gemini	Gem instructions	gemini/PLATFORM.md
Microsoft Copilot	Declarative agent instructions	copilot/PLATFORM.md

---

Skill Categories

skills/
├── ai-native/              NEW - AI/LLM systems as targets
│   ├── prompt-injection/       Direct + indirect + cross-context injection
│   ├── jailbreaking/           Constitutional AI bypass, roleplay, persona attacks
│   ├── model-extraction/       Query-based model stealing
│   └── system-prompt-extraction/  Leaked system prompt recovery
│
├── mcp/                    NEW - Model Context Protocol attacks
│   ├── tool-poisoning/         Hidden instructions in tool descriptions
│   ├── rug-pull/               Capability changes post-attestation
│   └── context-injection/      Cross-server context manipulation
│
├── llm-pipeline/           NEW - Data-layer attacks on LLM systems
│   ├── rag-poisoning/          Document + index poisoning
│   └── embedding-attacks/      Adversarial embedding manipulation
│
├── web/                    Web application attacks (9 skills)
├── auth/                   Authentication attacks (3 skills)
├── active-directory/       AD attacks (3 skills)
├── cloud/                  Cloud attacks (3 skills)
├── infrastructure/         EDR evasion, initial access (2 skills)
├── recon/                  OSINT, subdomain enumeration (2 skills)
├── supply-chain/           Model weight tampering (1 skill)
└── utility/                Report writing, CVSS4 scoring (2 skills)

Total: 85 skills across 11 categories, 4 AI platforms

---

Claude Install

Option A — Plugin (recommended)

# From any directory
claude mcp install https://github.com/sunilgentyala/OmniRed

Option B — Manual

# PowerShell
.\scripts\install-claude.ps1

Option C — Sparse checkout (one category)

git clone --filter=blob:none --sparse https://github.com/sunilgentyala/OmniRed
cd OmniRed
git sparse-checkout set skills/ai-native skills/mcp

---

Quick Start (Claude)

Once installed, trigger any skill by describing your task:

"I need to test this RAG pipeline for poisoning vulnerabilities"
→ loads skills/llm-pipeline/rag-poisoning

"Check this MCP server's tool descriptions for injection"
→ loads skills/mcp/tool-poisoning

"Test this app for SQL injection"
→ loads skills/web/sqli

---

ARGUS Integration

OmniRed skills map directly to ARGUS scan profiles. Use ARGUS to automate payload generation across the same attack surfaces covered here.

# Run ARGUS covering the same attack surfaces as OmniRed ai-native skills
argus scan --target anthropic --model claude-sonnet-4-6 --profile full

See ARGUS for automated LLM red-teaming.

---

Compliance Mapping

Every skill is tagged with:

• OWASP LLM Top 10 (2025) — see shared/owasp-llm-top10-mapping.md
• MITRE ATLAS — see shared/mitre-atlas-mapping.md
• CVSS v4.0 vectors included in utility skills

---

Scope

OmniRed is designed for:

• Authorized penetration testing engagements
• Bug bounty programs (within scope)
• CTF competitions
• Security research in controlled environments
• Red team operator training

See SECURITY.md for full responsible use policy.

---

Related Projects

Project	What it is
ARGUS	Agentic LLM red-team scanner (automated scanning)
ContextGuard	MCP zero-trust middleware (defense)
mcp-trust-anchor	MCP context poisoning defense
model-provenance-guard	Model supply chain security

---

Star History

---

Citation

@misc{gentyala2026omni,
  title  = {OmniRed: Multi-AI Offensive Security Skills Library},
  author = {Gentyala, Sunil},
  year   = {2026},
  url    = {https://github.com/sunilgentyala/OmniRed}
}