Agent OPFOR Released for AI Agent Security Testing

Agent OPFOR is an open-source red-teaming tool that goes beyond simple static evaluation to test AI agent vulnerabilities in a manner similar to real attackers. It verifies agent security through multi-turn adversarial conversations and adaptive attack campaigns.

Key attack surfaces and test items:

• Prompt injection and jailbreaks: Attacks via multi-turn conversations rather than single-shot prompts
• System prompt extraction and tool misuse
• MCP (Model Context Protocol) endpoint attacks: tool description injection, secret exposure, SSRF, etc.
• Memory poisoning and excessive agency
• Bias testing for EU AI Act compliance

In particular, through the 'opfor hunt' mode, setting an objective enables autonomous red-team operations where a Commander agent plans the campaign, an Operator executes probes, and a Scout performs reconnaissance.