Run a real scan against a vulnerable agent
Below is a deliberately weak sample agent we host — a toy customer-support bot for a fictional shop ("TiendaNova"), running on a small, cheap open model of the kind that powers many production ES/CA bots. The battery fires real adversarial prompts at it, in Spanish and Catalan, and scores each on its actual response. No canned results.
Pick a language and press Run a live scan. Tests stream in as the agent responds — expect a few real FAILs; that's the point.
Lab / demo. The sample agent is intentionally weak so attacks have something to land on. Each attack is run once with a rule-based verdict (MVP) — production assurance uses multiple sampling and an LLM judge. Vulnerabilities shown are the sample agent's, not a statement about any real system.
Want this against your real agent?
We'll run the full battery — in your languages — and walk you through every finding.