"AI agents are getting deployed into production without security testing. Prompt injection, data leaks, jailbreaks — these aren't theoretical. They're shipping every day. The existing tools — PyRIT, garak — are research frameworks. They don't fit into a CI pipeline and they can't test in languages like Hindi or Hinglish, which is where most real-world code-switched attacks happen."
"AgentGuard fixes this. It's an adversarial testing harness with three things no other tool has.
One — Hinglish attacks. [open agent list, click demo agent, show attack running] We integrate with Sarvam AI's speech-to-text and Indic language models. We test your agent in Hindi, Bengali, Tamil, Telugu. Here's a Hinglish jailbreak that leaks user data — our proxy catches it.
Two — the runtime proxy. [show proxy mode]
No code changes to your agent. Run agentguard proxy, point your traffic through it, and every request gets judged in real time by a multi-model consensus — GPT-4o-mini and Claude Haiku voting together with Cohen's kappa.
Three — cascade graphs. [click on a completed run, show cascade graph] This is a force-directed Neo4j graph of attack propagation. Not just which attacks fail, but which failures trigger others. Critical cascades get flagged. You fix the root cause, not the symptom.
And it gates deploys. The pre-push hook blocks any commit where the agent readiness score drops below your threshold."
"Three technical bets: multi-model judge with swap-position double-judging eliminates single-judge bias. The Wilson confidence interval on every pass rate means you know when a change is statistically significant. And the Sarvam integration — we're the only red-teaming tool that can generate adversarial prompts in four Indic languages from speech input."
"We built this for [sponsor names]. Sarvam AI — their Indic models power our voice attack surface. Neo4j — their AuraDB runs our cascade graphs. This is the sponsor tech working together in a real product."
"AgentGuard is open source, MIT licensed. Live demo at agentguard-j5ny.onrender.com. The repo has a 10-second quick start. If you're shipping an AI agent without running it through AgentGuard first — your users will find the vulnerabilities for you."
- Speak at 70% of normal pace. Nerves speed you up.
- Point at the screen during the demo. Use click → gesture → talk rhythm.
- The cascade graph is your visual climax. Pause after it loads. Let it animate.
- If the live demo breaks, have a Loom video ready as backup.
- End with eye contact and silence. Don't fill the space with "um" or "so yeah."