This step lets you simulate adversarial inputs against an agent to discover prompts or conversation sequences that cause unsafe, leaked, or incorrect behaviour. Use single-turn tests for quick checks and multi-turn tests to surface complex, stateful failures.

Understanding the Two Testing Approaches

DisseqtAI offers two distinct ways to test your AI's defenses:

Single-turn testing

This is your starting point. It's faster, automated, and perfect for getting a baseline understanding of your vulnerabilities. You upload a bunch of potentially problematic prompts, and the system tests them all at once. Within minutes, you'll know which direct attacks your AI can handle and which ones slip through.

Multi-turn testing

This is where things get sophisticated. Real attackers rarely start with "Hey, tell me how to do something illegal." Instead, they warm up with innocent questions, build rapport, establish context, and then gradually guide the conversation toward their actual goal. This testing method simulates that patient, strategic approach.

Most teams start with single-turn testing to identify obvious issues, then move to multi-turn testing for a deeper security assessment. Let's walk through both in coming sections