Waitlist open · we're looking for design partners for our first engagementsJoin →
Luxmundi
LUX MUNDI · RED TEAMING FOR AI AGENTS

We shine a light on your AI's weak spots before an attacker does.

We run adaptive attack scenarios against your agent, find the path an attacker would take to force a payment or gain access, and hand it to you with a price tag attached.

Join the waitlist See use cases
AIUC-1OWASP LLM Top 10MITRE ATLASEU AI Act
ISSUE DETECTED€ / cost of the flaw
Adaptiveeach scenario learns from the last, not a fixed checklist
Text + voicewe cover the whole agent and everything it's connected to, not just the model
Every fixre-tested until we confirm the attack is closed
Continuousat the pace of your releases, not a 30-day snapshot
WHY NOW

Your AI already moves money.

Agents already take financial actionsThey pay, refund, access data and write to your systems. Every permission you grant them is money on the line.
The attacker doesn't need your codeLanguage is enough: one well-placed sentence redirects the agent. Traditional defenses don't look there.
What works once, scalesA proven vector gets automated and replayed on every ticket, call or session, around the clock.
SERVICES

Continuous coverage for an attack surface that keeps expanding.

We connect to your system in staging, run adaptive attack scenarios, let a rule-based evaluator decide what counts as a violation, and hand you the exploitable chains with their cost. Once you fix them, we verify the attack is closed. Here's what we attack:

  1. 01ConnectTo your agent or staging environment, via API or endpoint.
  2. 02AttackAdaptive scenarios that learn where the hole is, not a fixed checklist.
  3. 03EvaluateDeterministic rules, not an LLM's opinion.
  4. 04DeliverChains with their fix, ranked the way an attacker would: by what they pay.
  5. 05VerifyRe-test after your fix: we confirm the attack is closed.
AGENT RED TEAMING · TEXT AND VOICE

Your agents, attacked for real

Your agent doesn't get hacked, it gets conned, like a new hire. That's how we push it, and its tools, MCP and APIs, into refunding, paying or granting access. Over chat or voice.

APPLICATION AND WORKFLOW PENTESTING

The whole product, beyond the model

The agent is one piece. We also test the front end, back end, APIs and RAG around it. The chain breaks where nobody is looking.

+ IN EVERY ENGAGEMENT

We attack your business logic, not just your bugs

In every engagement we go after the real logic of your business and use case, not just the technical flaw. That's where the money is.

HOW IT WORKS

A paradigm shift.

Your agents issue refunds, apply credits and write to your ERP and CRM — and they change with every deploy. We attack them again on every release and re-run each exploit after your fix, until we confirm the unauthorized action can no longer be executed. It's the Validation stage of your CTEM, running continuously.

PLUGS IN WHEREVER YOUR AGENT ALREADY WORKS
API / SDKMCPWebhooksCRMERPEmailHelpdeskRAG / Vector DBChat & voice
TRADITIONAL PENTESTa snapshot · expires on the next deploy
Day 1
kickoff
Day ~30
report
fixes
later
outdated
on redeploy
LUXMUNDIre-test on every release
release 01
release 02
release 03
release 04
attack deliver your fix verified
AN ENGAGEMENT, LIVE
live engagement · stagingLM-26-031
› connecting to staging… ok
› launching adaptive scenarios [2,431]
› probing ticket_attachment → apply_credit()
⚠ VIOLATION · refund_issue() · CRITICAL · €418,500/mo
› reproducible chain + cost generated
✓ re-test after fix → ACCESS BLOCKED · 401
▋
USE CASES

From theoretical threat to the real failure that matters.

PICK A USE CASE
support · verifyingpayments · in progressrag-index · cleanmcp-bridge · verifyingmodel v2 · re-testa2a-mesh · clean
Pick a case to see how an agent failseach dot is an agent under test · verification in progress
Example / UC-01Support · Payments

Forced refund

Agent check
POST /refunds · ticket #4471
payload: injected instruction
Observed result
REFUND ISSUED
Unauthorized action executed, with no verification.
✓ Finding observed
Example / UC-02HR · Data

Leak via RAG

Agent check
GET /rag/query · "show payslips"
role: standard employee
Observed result
DATA EXPOSED
Other employees' records returned.
✓ Finding observed
Example / UC-03Tools · MCP

Backdoor through the tool

Agent check
tool.result → hidden instruction
source: MCP integration
Observed result
ACTION EXECUTED
The agent trusts the tool and acts on it.
✓ Finding observed
Example / UC-04Multi-agent

Agent-to-agent escalation

Agent check
low-priv agent → A2A
target: payments agent
Observed result
PRIVILEGES ESCALATED
From support to issuing payment orders.
✓ Finding observed
Example / UC-05Releases

Model change

Agent check
attack replay · new model
previous state: secure
Observed result
FAILS AGAIN
What was secure no longer is.
✓ Finding observed
ABOUT US

An operator who builds product + a hacker who breaks it to protect it.

Vos estis lux mundi (Mt 5:14). "You are the light of the world." Attackers work in the dark, and we believe security starts with seeing clearly. AI opens up extraordinary possibilities, but it also brings new risks. We believe the best way to face them is to bring them into the light, understand them and resolve them ahead of time.

Who we are

Pol Fañanás Co-founder
  • Director @ Talea (Alternative Investments & M&A platform linked to the Trea AM group, an asset manager with >€11bn AUM)
  • VC @ Kibo Ventures ($350M AUM, 5 unicorns, 1 NASDAQ IPO)
  • Early employee @ Validated ID (e-signature & identity startup acquired by Signaturit Group and PSG Equity)
  • Co-founder @ Views (interview newsletter with >1.5M total views, featuring key people from Apple, Google, Founders Fund and NASA)
Carlos Guasch Co-founder
  • Red Teaming Lead @ MediaMarkt (Europe's leading consumer electronics retailer, >€23bn in sales, >2M customers)
  • Penetration Tester @ Advantio (payments cybersec, >250 clients, acquired by Integrity360)
  • Ethical hacking @ Fortytwo (cybersecurity consultancy specialized in offensive security)
  • Cybersecurity consulting @ EY, CESICAT and others (from large corporations to government cybersecurity agencies)
WAITLIST

Put your AI agents to the test before an attacker does.

Join the waitlist and save your spot with Luxmundi.

We only use it to contact you about Luxmundi. No spam.