KeyValueSoftwareSystems/agent-opfor

What it solves

Opfor is an adversary emulation tool designed to red-team AI agents, LLM applications, and MCP servers. It helps developers and security teams identify vulnerabilities such as prompt injection, jailbreaks, tool misuse, and sensitive data disclosure before they are deployed to production.

How it works

Opfor simulates real-world attacks by generating targeted adversarial prompts based on industry standards (like OWASP). It then executes these attacks against a target agent, captures the responses, and uses an LLM-based judge to classify each attempt as a pass or fail with detailed reasoning. The tool can also integrate with observability stacks (like Langfuse or Netra) to analyze internal tool calls and reasoning steps rather than just the final output.

Who it’s for

• Engineers and DevOps: Use the CLI or SDK to integrate red-teaming into CI/CD pipelines.
• Product Managers and QA: Use the browser extension to test deployed chatbots without writing code.
• Security Analysts: Use the curated OWASP suites to ensure compliance and safety.
• AI Developers: Use the MCP server mode to let their coding agents test other agents.

Highlights

• Comprehensive Coverage: Includes curated suites for OWASP LLM Top 10, Agentic AI Top 10, MCP Top 10, API Security, and EU AI Act bias.
• Multiple Entry Points: Available as a CLI, browser extension, MCP server, IDE skills, and a TypeScript SDK.
• ** uma Autonomous Red-Teaming**: The opfor hunt command uses a multi-agent system (commander, operators, scout) to autonomously conduct recon and adaptive attack campaigns.
• Trace-Aware Testing: Integrates with telemetry providers to detect leaks that occur in internal tool calls but don't appear in the final response.
• Transparent Reporting: Generates HTML and JSON reports with full logs of every attack prompt and judge verdict, including a breakdown of the testing cost (token usage).