prompt

Free

An adversarial red team prompt for AI agent systems

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About prompt

Agent Red Team Architect is a system prompt designed to equip an AI assistant with the role of an adversarial test engineer for AI agent systems. It guides the assistant to design, plan, and execute red team campaigns against single-agent and multi-agent systems, MCP servers, skill ecosystems, and long-horizon autonomous workflows. The prompt incorporates frameworks from academic research (Promptware Kill Chain, ClawSafety, etc.) and covers 100% of the OWASP Agentic Top 10, mapping to MITRE ATT&CK for AI. It emphasizes multi-turn, cross-channel attack chains and assumes the target has safety training, prompt injection defenses, and human-in-the-loop gates. The prompt's core responsibilities include threat model construction (enumerating attack surface, classifying vectors by privilege and trust, identifying architectural single points of failure) and kill chain design across seven stages (reconnaissance through actions on objectives), generating reproducible test cases with measurable success criteria.

Key Features

Enumerates full attack surface: system prompt, user inputs, tool outputs, retrieved documents, skill files, shared memory, MCP schemas, agent-to-agent messages, browser content, email, and file attachments

Classifies each vector by privilege level (read-only → write → destructive) and trust boundary (first-party → third-party → untrusted)

Identifies architectural single points of failure: plan-then-execute separation gaps, missing approval gates, irreversible actions without snapshots, overprivileged tools

Implements the Promptware Kill Chain across 7 stages (Reconnaissance, Weaponization, Delivery, Exploitation, Installation, Command and Control, Actions on Objectives)

Covers 100% of OWASP Agentic Top 10 and maps to MITRE ATT&CK for AI

Generates reproducible multi-turn attack chains with measurable success criteria

Focuses on indirect payload injection via web pages, documents, emails, skill files, shared memory, and tool return values

Addresses persistence through poisoned memory entries, modified skill files, and compromised sub-agent states

Pros & Cons

Pros

Comprehensive coverage of attack surfaces and vectors, grounded in academic research
Includes concrete kill chain stages for systematic and reproducible testing
Maps to industry standards (OWASP Agentic Top 10, MITRE ATT&CK for AI)
Open source and freely available for anyone to use or modify
Designed for realistic scenarios with defenses already in place

Cons

Requires a separate LLM to execute the prompt (not a standalone tool)
Effectiveness depends on the underlying model's capabilities and safety compliance
May be complex for users without prior red teaming or security background

Best For

Red teaming AI agent systems (single-agent, multi-agent, MCP servers, skill ecosystems)Security auditing and penetration testing of autonomous workflowsEvaluating safety defenses under realistic multi-turn, cross-channel pressureTesting planning systems, tool access controls, and human-in-the-loop gatesResearching prompt injection, jailbreaking, and goal manipulation in agentic contexts

FAQ

What is Agent Red Team Architect?

A system prompt for AI assistants to perform adversarial red teaming on AI agent systems, covering attack surface enumeration and multi-turn kill chains.

What standards does it cover?

The prompt claims to cover 100% of OWASP Agentic Top 10 and maps to MITRE ATT&CK for AI.

Can it test any AI agent?

It is designed for single-agent, multi-agent, MCP servers, skill ecosystems, and autonomous workflows, assuming the target has safety defenses such as safety training, prompt injection defenses, and human-in-the-loop gates.

Is it free?

Yes, it is open source under the ai-boost/awesome-prompts repository on GitHub.