Xtrinel builds VAAST, the vulnerability scanner for LLM-integrated applications. Find prompt injection, tool-call abuse, RAG poisoning, and agentic pipeline flaws before they reach production.
Supported by

The AI Attack Surface
LLM-integrated applications introduce a class of vulnerabilities that conventional scanners were never built to reach. VAAST covers the gap.
Adversarial inputs that hijack instructions, exfiltrate system prompts, or coerce the model into forbidden actions. VAAST catalogs direct, indirect, and multi-turn variants.
Agentic stacks expose tool surfaces as first-class APIs. VAAST's MCP Scanner enumerates callable tools and probes each one for parameter injection, privilege escalation, schema confusion, and unauthorized state changes.
Retrieval layers trust their corpus. VAAST inspects ingestion paths, checks for prompt-level poisoning, and flags documents that coerce downstream reasoning.
Chained agents, plan-and-execute loops, and autonomous workflows create new blast radii. VAAST maps the full graph and identifies failure modes unique to long-horizon runs.
VAAST in Action
Point VAAST at your LLM endpoint, run the payload catalog, and review and export what it finds.

Run VAAST's payload catalog against your LLM app.
Xtrinel
How VAAST is different
No black boxes. Every finding comes with evidence, every detection is documented, and VAAST only runs where you're authorized to test.
Detections are documented in public writeups from HYDRACUDA, our research division.
VAAST is built for practitioners with written authorization. Our license agreement requires you to say so before every engagement.
Every finding includes the exact payload and the model’s response, exportable as JSON or HTML.
Bring VAAST into MCP-compatible AI agents and clients. Read-only access to your local scan findings and workspaces.
npx -y @xtrinel/vaast-mcpStart with the free tier. Upgrade when your findings pile up.