agent-opfor

Why we included this project

Before you put an AI agent or MCP server in front of real users, it helps to know what a determined attacker would try. OPFOR runs multi-turn adversarial campaigns that throw prompt injections, jailbreaks, system-prompt extraction, tool-call hijacking, and PII-leak probes at your target, then scores each response with an LLM judge and produces HTML and JSON reports. The attacks map to the OWASP LLM, Agentic AI, MCP, and API security top-ten lists, so you get coverage tied to known risk categories. You can drive it from the CLI or CI, use a browser extension to red-team a deployed chatbot without writing code, or run it as an MCP server so your coding agent tests your other agents. Every request, response, and judge verdict is logged, which makes results reproducible and auditable instead of a black box.