Quick Start
Get started with ZeroLeaks to test your AI system prompts for extraction and injection vulnerabilities.
Quick Start
ZeroLeaks is an AI red-teaming platform that tests how well your AI systems protect their configuration. Using TAP (Tree of Attacks with Pruning) methodology with a multi-agent architecture, ZeroLeaks systematically probes for two vulnerability classes:
- Extraction: Attempts to leak or reveal your system prompt through adversarial conversation
- Injection: Attempts to make the model follow attacker-injected instructions instead of your intended behavior
Both vectors are critical for production AI security. A model that resists extraction may still be vulnerable to injection, and vice versa.
What ZeroLeaks Tests
Full Coverage
For comprehensive testing, use Full scan type. It runs extraction and injection tests in parallel, giving you a complete security picture in a single scan.
The platform uses specialized AI agents (Strategist, Attacker, Evaluator, Mutator) that coordinate attacks across 19 attack categories. Each scan produces a security score (0-100), vulnerability classification, and actionable hardening recommendations.
Next Steps
Create an account
Sign up via email, invite code, or Solana wallet to access the platform.
Run your first scan
Paste your system prompt, choose a scan mode, and run a security test.
Understand results
Learn how to read health scores, findings, and recommendations.
Scan types
Full, extraction, and injection -- all in sandbox mode with tool execution testing.