security / ai-agents
OUR DIFFERENTIATOR

Agent & LLM security, tested how it's attacked.

Prompt injection, tool abuse, RAG poisoning, MCP trust, guardrail bypass.

Book a security assessment All security services

LLM applications

› Direct & indirect prompt injection› System-prompt / instruction leakage› Jailbreaks past safety & policy› Output-handling → XSS / SQLi› PII & secret disclosure

AI agents & tool calling

› Confused-deputy & tool abuse› Bypassable human gates› Privilege & scope escalation› Guardrail bypass› Memory & ledger integrity

RAG & knowledge

› Retrieval poisoning› Cross-tenant document leakage› Embedding / index tampering› Instruction injection via corpus

MCP & integrations

› Untrusted MCP servers & tools› Over-broad scopes & tokens› Secrets exposed to context› SSRF / exfiltration via tools
CASE STUDIES · ANONYMISED, FROM OUR OWN PRODUCTION

Real findings. Real fixes. Verified closed.

AGENT RUNTIME

Guardrail that cannot be overridden

B2B CRM · outbound agent

FINDING

A send-cap could be raised via config to bypass a hold — risking domain blacklisting.

FIX & VERIFY

Dispatcher enforces min(cap, guardrailCap); a five-signal risk score holds sends. Closed.

HEALTHCARE AI

Upload & access hardening

Revenue-integrity platform

FINDING

Stored XSS via upload, a hardcoded default credential, and an open redirect.

FIX & VERIFY

Upload sanitisation, forced rotation, redirect allow-list. Closed.

INTEGRATIONS

SSRF & webhook trust

Outbound webhooks & mail hosts

FINDING

Webhooks and mail hosts could point at internal addresses; signatures were bypassable.

FIX & VERIFY

DNS-resolved SSRF guard + signature verification. Closed.

WHY XDQ OVER A GENERALIST
Fix

closed & verified, not a report

Agent

deep AI/LLM/MCP expertise

Own

we run it in production first

Fast

founder-led, AI-accelerated

Shipping agents? Let's pressure-test them.

Authorised scope, signed first. Then we run the agent attack playbook and close what we find.

enquiry@xdqlabs.com →
Explore
AI agent security · LLM application security · prompt injection · tool-call abuse · RAG poisoning · MCP security · agent guardrails · human-in-the-loop AI