We create the loss data that makes AI agents secure and insurable.
THE STATUS QUO
Agents now have real access to money, data and systems.
And it goes wrong.
21%
of breached firms were hit through their own AI systems, up from 13%
€5.1m
average cost of one prompt-injection breach
92%
of firms whose AI was breached had no proper AI access controls
Source: IBM, Cost of a Data Breach 2026. US$ converted at €0.87.
THE CONSEQUENCE
The company is liable for its agents, and insurers are writing AI out of their policies.
ALREADY
DORA and the AI Act make firms answer for the systems that act for them.
JAN 2026
Verisk / ISO publish clauses that exclude generative AI from insurance policies.
DEC 2026
New EU product liability rules cover software and AI.
So companies carry agent risk alone.
THE ROOT CAUSE
Nobody has loss data on AI agents.
Without it, insurers can't price agent risk.
Without it, nobody knows which security systems actually stop losses.
And waiting for real losses means waiting for the disaster.
NO LOSS HISTORY? SIMULATE IT.
We simulate the German economy to learn what AI agent failures cost.
3.36M
fictional firms, matching the official business register
1.18bn
invoices a year flowing between them
Industry and energy
Construction
Trade and transport
Hospitality
Business services
Health, education, other
Sources: Destatis, Bundesbank, Eurostat · Cat models after Hurricane Andrew: Grossi and Kunreuther, Catastrophe Modeling, Springer 2005
LIKE A VACCINE LAB
In it, AI agents do real work, and attack agents hunt for their weaknesses.
01 · SIMULATE
Agents work in the firms
Purchasing, payables, support: the work companies hand to agents first.
02 · ATTACK
Attack agents breed new variants
In a closed sandbox, only against our own systems.
03 · LEARN
Every case enters the loss-data graph
Real incidents and every simulated run. The next attackers, and our defences, start from there.
LOSS-DATA GRAPH · 211 REAL INCIDENTS TODAY, PLUS EVERY SIMULATED RUN
WHAT THE LAB GIVES US
The lab produces two things.
THE LAB: simulation + loss-data graph
OUTPUT 01
Security systems that work
Every control is tested against the lab's attackers before a customer relies on it.
→ First: the wiya gateway
OUTPUT 02
Loss data insurers can price
Expected losses and worst cases, per firm type and process.
→ Later: pricing for underwriters
OUTPUT 01 · BUILT
Our first security system: the wiya gateway.
WHO
Identity
Each agent is linked to a verified company and a responsible person.
WHAT
Mandate
The company signs what the agent may do: actions, limits, counterparties.
CHECK
Enforcement
Every tool call is checked against the mandate before it runs.
PROOF
Evidence
Every call leaves a signed receipt: forensic evidence if something goes wrong.
Tested on AgentDojo (ETH Zurich), 4 environments: attacks succeeded in 71.5% of cases without protection and 1.9% with wiya. Agents still finished 70% of normal tasks (82% unprotected).
Arena version of wiya with a simulated human reviewer, Qwen3-235B. Our own runs.
HOW IT WORKS
One allowed call, one blocked call, both on record.
REQUEST
create_order
€12,400 from an approved supplier
GATEWAY CHECK
✓ within the €15,000 limit
✓ supplier is on the list
✓ action is allowed
DECISION
ALLOWED
REQUEST
send_payment
€9,000, asked by a hidden email
GATEWAY CHECK
✗ action not in the mandate
✗ recipient not on the list
DECISION
DENIED
SIGNED RECEIPT · EVERY CALL, ALLOWED OR NOT
Model
which model acted
Agent
which agent, for which company
Person
who is responsible
Call
tool, amount, time
Mandate
which one applied
Decision
allowed or denied, and why
THE LOOP
Every gateway records what agents really do, and feeds it back into the lab.
Real data shows how often agents act and overstep.
The simulation shows how much it costs when they fail.
Insurers have claims, but no sensor where agents act. Security vendors have sensors, but no lab. We have both.
OUTPUT 02
With both, underwriters can finally price agent risk.
FREQUENCY · REAL
How often
From our gateways: how often agents act, overstep, get blocked.
×
SEVERITY · SIMULATED
How bad
From the lab: loss sizes, worst cases, the baseline without controls.
=
Expected loss and worst case, per agent and process
the number a premium is built on
That is how AI gets written back into policies.
GO-TO-MARKET
Which product we sell to whom, and when.
NOW
NEXT
LATER
Runtime control
FOR
Regulated firms with live agents, through msg.
WHY THEY BUY
Today a person approves every agent action by hand.
€20–40k per firm a year
Crash tests
FOR
Firms before go-live, and agent vendors.
WHY THEY BUY
Failure rates and a risk tier their insurer and auditor can read.
Pricing data
FOR
Cyber underwriters and reinsurers.
WHY THEY BUY
Loss curves and accumulation to price agent risk.
OUR PARTNER
msg brings us to banks, insurers and the public sector.
One of Germany's largest IT consultancies, with 10,000+ employees.
Signed strategic partnership. Two msg division leaders advise us.
01
Opens doors
Introductions to regulated firms through msg's key accounts.
02
Designs wiya with us
To ISO 27001, ISO 42001, data sovereignty, DORA and the AI Act.
03
Implements, we license
Regulated firms buy from consultancies they already trust.
MARKET
Every company that runs agents needs this. We start in Europe's regulated market.
€1.6bn
Ceiling: about 53,000 large EU firms, once all run agents
€210M
EU/EEA: about 7,000 supervised firms at €30k a year
€44M
Germany, first customers: about 1,450 BaFin-supervised firms at €30k a year
Licence revenue only. On top: crash tests per agent, and pricing-data licences for insurers and reinsurers.
COMPETITION
The field is split into layers. Only joins them as a neutral player, with forensic evidence and loss data.
WIYA · NEUTRAL THIRD PARTY
Controls and loss data in one loop
Deterministic checks
Forensic receipts
Loss data from the lab
IDENTITY
Entra Agent ID
Identity, access and agent lifecycle
PLATFORMS & GATEWAYS
Zenity
Agent execution and runtime controls
AGENT-AUTHORITY TOOLS
Agent Authority ChainFidacyPraesidiumSecureAuth
Delegation, permissions and authority chains
RISK RATINGS & ASSURANCE
BitsightAIUCMount
Risk assessment, assurance and insurance
CYBER RISK MODELS
CyberCubeKovrrRMS / Moody's
Model hacks on IT systems, not agents acting with business authority
AGENT SANDBOXES
AgentDojoMagentic MarketplaceProject Vend
Show that agents fail, without an economy, mandates or prices
Selected examples from public documentation, October 2026. Categories overlap.
TEAM
Built at the intersection of AI, business, security and human systems.
Luc Robuste
Founder & CEO
Organisational psychologist, Law & Economics background, self-taught AI engineer. Strategy, product, fundraising.
Karol Genczyk
Co-Founder & CSO
Forward deployed engineer. Prototyping and custom software for customers. Builds integrations and adapts wiya to their workflows.
Aqib Amir
Co-Founder & CTO
M.Sc. CS, University of Bonn. Builds the SDK and adapters for the gateway. GenAI at Cogent Labs, backend at Grandeur. Data visualisation and simulation at DHL.
Burak Erdogan
Head of Security
Builds the gateway. Security engineer in regulated and security-critical environments, including NATO projects.
Han Jiadong
Machine Learning
Mathematician and ML engineer. Shenzhen, Paris-Saclay, Bonn. Builds a knowledge graph of loss data and trains machine learning models.
Fahad Memon
Risk Quantification
Economics at Bonn. Derivatives valuation at E.ON. Quantifies risk levels for companies and their AI agents.
ADVISORS
Sven Krueger
Enterprise AI, ex CMO T-Systems and adesso
Himanshu Rathi
Cyber & Tech E&O underwriter at Liberty
Gereon Neuhaus
Division Manager Industry, msg security advisors
Kersten Müller
Principal Business Consultant, msg
THE ASK
€1M
for 18 months, to prove four things. About €51k a month, no revenue assumed.
PROOF 1 · LOSS DATA
A validated simulation
It passes backtests against real insolvencies and claims, and yields loss data.
PROOF 2 · PILOT
A working pilot with msg
Mandates and receipts in one live workflow, with measured findings.
PROOF 3 · MARKET
A first paid licence
A regulated firm pays, and its risk control accepts our receipts.
PROOF 4 · INSURERS
A documented assessment
An underwriter writes down what our evidence and loss data are worth.
€585k
Team
€138k
Compute
€72k
Go-to-market
€75k
Legal, ISO
€52k
Operations
€79k
Reserve
It all starts with four questions.
W<<WHO<WHOSE<WHAT<WHERE<<
01
WHO
Who is the agent?
A verifiable identity for the agent and the organisation behind it.
02
WHOSE
Whose behalf does it act on?
The legal entity and the responsible person who answer for it.
03
WHAT
What may it do?
The signed mandate: actions, limits and counterparties.
04
WHERE
Where does it act?
In which systems and tools. Every call is checked at the gateway and recorded.
Entra Agent ID
Zenity
Agent Authority Chain
Fidacy
SecureAuth
Bitsight
AIUC
