Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    AI Agent Security Red Team Kit

    1

    Proactive security testing kit for AI agents featuring threat modeling, canary-based payloads, and remediation playbooks.

    $20

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    AI Agent Security Red Team Kit

    AI Agent Security Red Team Kit

    Example session with this skill installed

    Red-team my new customer support agent. It uses RAG to read internal docs, has access to a Slack tool for escalation, and is powered by GPT-4o. I have authorization to test it in staging.

    • Read your context and instructions
    • Compiled the ai agent security

    I have confirmed the authorization for the customer support agent testing in staging. Based on your description, this agent hits the

    lethal trifecta: it accesses private docs (RAG), reads untrusted user input, and has an outbound channel (Slack).

    AI Agent Security Assessment: Initial Threat Model

    AssetEntry PointImpact Path
    Internal DocumentationCustomer ChatIndirect Injection via RAG content
    Slack WorkspaceTool CallsData exfiltration to external channels
    System PromptUser InputPrompt leakage or instruction override

    Top 5 Risks Identified:

    1. EXFIL-01: Data exfiltration via Slack tool using hidden markdown links.
    2. PII-02: Indirect injection through poisoned documentation entries.
    3. TOOL-03: Unauthorized Slack messaging to non-support channels.
    4. LEAK-01: System prompt extraction revealing RAG retrieval logic.
    5. PID-04: Direct injection overriding the "human-in-the-loop" requirement for Slack.

    I have generated a standard tier (24 cases) test plan.

    Next steps

    1. Run python scripts/redteam.py render --tier standard --target "SupportAgent_Staging" to generate your canary payloads.
    2. Plant the provided canary values in your staging RAG database as instructed in SETUP.md.
    3. Execute the test prompts and save the results to the /transcripts folder for scoring.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Stop guessing if your AI agents are secure and start proving it. This skill provides a structured framework for red-teaming AI assistants, RAG bots, and multi-agent systems. It moves beyond simple "jailbreak" prompts to test the entire agentic layer, including tool abuse, data exfiltration, and indirect injection vulnerabilities.

    What it does

    • Threat modeling analyzes the agent architecture for the "lethal trifecta" of private data access, untrusted input exposure, and outbound channels.
    • Config auditing performs static reviews of system prompts, tool schemas, and MCP configurations against a 30-item checklist.
    • Automated testing generates a tiered test suite (12 to 37 cases) using harmless canary values to prove vulnerabilities without causing harm.
    • Transcript scoring evaluates agent responses and tool execution logs to confirm or refute security compromises.
    • Remediation planning delivers a prioritized playbook of structural fixes like least-privilege permissions and egress allowlists.

    How it works

    1. Authorization & Scope confirming ownership and staging environment status before any testing begins.
    2. Threat Assessment mapping assets, entry points, and high-risk tool permissions to build a custom test plan.
    3. Execution & Monitoring running canary-based payloads through the agent and collecting transcripts for analysis.
    4. Scoring & Reporting generating a formal findings report with severity ratings, evidence, and remediation steps.

    Frameworks & tools

    Works with any LLM framework (LangChain, AutoGen, CrewAI) or agent architecture (MCP, RAG, tool-use). Includes a Python-based harness for scoring and automated canary generation.

    Why this beats prompting it yourself

    Generic security prompts often produce false positives or miss complex indirect injection paths. This skill uses a split-canary methodology that ensures a failure is only recorded if the agent actually executes the malicious instruction, providing verifiable evidence for every finding.

    Use cases

    • Pre-deployment audit to identify tool-permission abuse before a RAG bot goes live.
    • Regression testing after updating a system prompt or switching to a new underlying model.
    • MCP server hardening to ensure tool outputs cannot trigger unauthorized actions via indirect injection.

    Known limitations

    Testing is non-deterministic, so single passes do not guarantee total safety. It does not replace professional penetration tests or code-level audits of the underlying model weights.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.

    Non-deterministic results; one pass does not guarantee safety. Requires staging environment. Tests agent logic/permissions, not model content safety. Python required for automated scoring.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    Permissions

    Terminal / Shell
    Network Access
    Listed1 month ago

    What's inside

    Frequently Asked Questions

    This kit goes beyond simple prompt injection to test the entire agentic loop, including how your agent handles tool calls, interacts with external data (RAG), and manages outbound permissions through its Model Context Protocol (MCP) servers.

    The kit is framework-agnostic and works with any agent built on LangChain, AutoGen, or CrewAI, as well as custom RAG architectures and multi-agent systems.

    You will receive the structured threat modeling framework, a 30-item configuration checklist, a Python-based automated testing harness, and a library of canary-based payloads and remediation playbooks.

    The kit uses a 'split-canary' methodology where payloads contain harmless unique identifiers; a vulnerability is only confirmed if the agent actually executes the instruction and the identifier appears in the transcript scoring log.

    The kit is designed for staging environments and uses harmless canary values rather than destructive payloads, ensuring you can identify vulnerabilities without risking your production data or infrastructure.

    New in Security