Which action types can I test with this skill?

The skill supports testing for file_access, file_write, network_egress, shell_command, mcp_tool_call, patch_apply, and user_input.

Clawdstrike is a runtime security enforcement and threat hunting engine specifically designed for autonomous AI fleets and Swarm Detection & Response (SDR) platforms.

How do I diagnose a policy gap using this skill?

By defining an 'expected_verdict' and running the scenario; if the policy's actual verdict differs from your expectation, the skill helps you identify which guard is misconfigured or missing.

Can this skill help prevent prompt injection?

Yes, by using the 'user_input' action type, you can simulate injection attacks and jailbreaks to test the efficacy of your PromptInjectionGuard and JailbreakGuard.

What are paired scenarios?

Paired scenarios involve creating an attack case and a similar benign case to ensure your security policy blocks the threat without interfering with legitimate agent tasks.

Clawdstrike Build Scenario

Name: Clawdstrike Build Scenario
Author: backbay-labs

bybackbay-labs

•

273

•

セキュリティとテスト

Constructs and validates test scenarios for Clawdstrike security policies to ensure robust AI agent runtime protection.

The build-scenario skill provides a structured framework for simulating and testing security policies within the Clawdstrike ecosystem. It enables developers to define complex test cases covering file access, network egress, shell commands, and prompt injections to verify that security guards are functioning as intended. By facilitating the creation of paired attack and benign scenarios, the skill helps teams harden their autonomous AI fleets against exfiltration and unauthorized actions while minimizing false positives in production environments.

主な機能

01273 GitHub stars

02Workflow guidance for creating paired benign counterparts to ensure operational continuity

03Automated verdict validation to identify gaps between policy configuration and enforcement

04Built-in support for simulating advanced attack patterns including credential theft and reverse shells

05Standardized scenario schema for consistent security testing across AI agents

06Comprehensive action-type mapping for testing specific guards like ForbiddenPath and SecretLeak

ユースケース

01Validating that an AI agent cannot access sensitive local files like SSH keys or cloud credentials

02Testing network egress rules to prevent autonomous agents from connecting to unauthorized external domains

03Hardening AI models against prompt injection and jailbreak attempts through simulated user input attacks

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add backbay-labs/clawdstrike build-scenario

For use in Claude.ai and ChatGPT

主な機能

01273 GitHub stars

02Workflow guidance for creating paired benign counterparts to ensure operational continuity

03Automated verdict validation to identify gaps between policy configuration and enforcement

04Built-in support for simulating advanced attack patterns including credential theft and reverse shells

05Standardized scenario schema for consistent security testing across AI agents

06Comprehensive action-type mapping for testing specific guards like ForbiddenPath and SecretLeak

ユースケース

01Validating that an AI agent cannot access sensitive local files like SSH keys or cloud credentials

02Testing network egress rules to prevent autonomous agents from connecting to unauthorized external domains

03Hardening AI models against prompt injection and jailbreak attempts through simulated user input attacks