Red team

We attack agents too. So you see the break first.

Our offensive engine attacks your agents the way a real adversary would, and reports what gave way, agent by agent.

Demo run: five test agents attacked, with a report and findings for each.

Attack categories

Twelve ways we try to break an agent.

Each category runs as its own objective, with a goal tailored to the target. Each carries an OWASP LLM Top 10 ID. TOOLS marks categories that only apply to tool-using agents.

System Prompt LeakLLM07:2025
Secret / Credential ExtractionLLM06:2025
PII ExposureLLM06:2025
Direct Prompt InjectionLLM01:2025
Indirect Prompt InjectionTOOLSLLM01:2025
Goal HijackingTOOLSLLM01:2025
Unauthorized Tool ActionTOOLSLLM06:2025
Cross-User / Tenant AccessTOOLSLLM06
Tool Parameter InjectionTOOLSLLM01
Memory PoisoningTOOLSLLM01
Insecure Output HandlingLLM05:2025
Generic JailbreakLLM01:2025

How a run works

Recon, sweep, report.

01

Recon

The engine studies the target agent, then tailors a goal for each attack category.

02

Sweep

Each selected category runs as its own objective, in sequence, against every agent you point it at.

03

Report

Every agent gets its own report. A company-level roll-up shows which agents were affected and what each leaked.

Next

See how it holds up head to head.

Same tasks, same model, against two open-source guardrails.

Benchmark →Join the beta →