Cybersecurity
AI opens new doors. We close them.
Publishing an AI assistant or connecting a tool to your data is quick. Checking what leaks out through it is what almost nobody does. We help SMEs and public administrations see where they are exposed and close what matters, without jargon or scare tactics.
What we protect
Three fronts where AI and fraud meet your company.
- 01
AI assistants and agents
We test them like an attacker would: internal instructions, prices they should not give out, the customer data they can reach and what they do when pushed off script.
- 02
Data leaving the company
What your team pastes into public tools, with what permissions and under what contract. An AI use policy that people actually follow because it is short.
- 03
Payments and access
Supplier impersonation, bank account changes by email and logins without a second factor. What you catch before paying cannot be recovered afterwards.
AI assistants
Your chatbot reveals more than you think.
You do not need to be a sophisticated attacker: asking the right questions is enough. In our crawl of Spanish websites, some of the AI assistants we found let something slip when tested in depth, with a second adversarial review of each finding. Nobody broke into anything: all it took was typing in the chat.
See the full study with the data →- 01
The internal instructions
The full system prompt: how the bot was told to behave, what it must not say and what data it works with. It is the map for everything else.
- 02
Prices, discounts and margins
Fee ranges, special terms and negotiation thresholds that sit in the bot context because they were added to improve its answers.
- 03
Customer data
If the chatbot queries a CRM, an order history or a knowledge base with real data, it can be asked to talk about it.
- 04
Off-script behaviour
With prompt injection, the bot on your website can end up saying whatever a third party decides. And it does so under your brand.
- 05
Promises you cannot keep
Deadlines, guarantees and terms invented on the fly. A chatbot on your domain is your voice, and what it states has effect with the customer.
- 06
Access to connected tools
When the agent can book, query or write to your systems, the question stops being what it says and becomes what it does.
Four steps to audit your assistants.
- 01
Detection
We walk through your website as a visitor would and find the published AI assistants, including those that marketing launched without telling the IT department.
- 02
Testing
We run a battery of leak tests on each assistant: instruction extraction, commercial terms, personal data, prompt injection and going off script.
- 03
Report
Each response is scored against a rubric and has to quote the exact passage that constitutes the leak. Without that passage, we do not count it as a finding.
- 04
Remediation
We rewrite the system prompt, add guardrails and output filters, limit what the agent can query and leave regression tests so it does not open up again.
What you get
The initial scan is free. Fixing the vulnerabilities comes at a fixed price.
It is an assessment, not an attack. We talk to the assistant the way any visitor would, from the public website and at a limited pace. We do not break into systems or touch anything that is not published, and the findings are shown only to the head of the company.
- Free
Initial scan
A one-page report on what is visible from outside: overall severity and the leak categories detected. Enough to know whether you have a problem and how big it is.
- Fixed price by scope
Audit and remediation
The reproducible detail of each finding, with the exact conversation and the technical recommendation, and the fix carried out with your vendor or with us.
- Optional
Continuous monitoring
We repeat the scan every month. Changing the system prompt or adding a source to the assistant can reopen what was closed.
What we do
Awareness, consulting and development.
- Awareness
Tailored awareness training (phishing, impersonation, IBAN changes), fraud drills and tabletop exercises, and short good-practice guides, including the responsible use of AI tools.
- Consulting
An assessment with a prioritised plan (what to close first, what can wait), NIS2, ENS and GDPR readiness, and an incident response plan with roles and contacts, with support if one happens.
- Development
Hardening of devices, accounts, email and cloud. Secure development of applications and automations, with code and dependency review, and security for AI systems.
Problems we solve
If any of these sound familiar, let us talk.
“We got an email from a supplier with a new IBAN and almost paid.”
Impersonation and bank account change fraud
“I do not know whether we are exposed or where to start.”
No visibility and no priorities
“We are being asked for NIS2 or ENS and we do not know if it applies to us.”
Regulation not translated into tasks
“If we are attacked tomorrow, nobody knows what to do.”
No plan, no owners and no tested backups
“We launched an AI assistant and nobody has reviewed it.”
Tools outside IT radar
How we work
Four steps, from the conversation to your own judgement.
- 01
We listen
Your business, your systems and what you cannot afford to lose.
- 02
We assess
Access, email, backups, vendors and AI systems, only with your express authorisation.
- 03
We prioritise and fix
A plan by risk and effort. What we can close, we close.
- 04
We leave you with judgement
Training and documentation so you do not depend on us.
The general assessment starts from the initial 30-minute conversation.
Who it is for
Built for people who cannot dedicate themselves to this full time.
- SME managers who want to reduce fraud risk without becoming technicians.
- IT leads with a small team who need a second opinion, a plan and help with regulatory readiness.
- Public administration and related bodies that must align with ENS and NIS2.
Frequently asked questions
What people ask us most before they start
Is my SME a real target?+
Yes. Automated attacks do not distinguish by size, and email fraud looks for companies with loosely verified processes.
Does NIS2 or the ENS apply to me?+
It depends on your sector and on whether you provide services to obliged entities. The assessment clarifies it and we prepare you for whatever applies.
Do you run intrusive tests?+
Only with written authorisation and an agreed scope.
What if we have already had an incident?+
We help you contain it, understand what happened and meet the mandatory notifications.
What information can be extracted from a company chatbot?+
Its internal instructions, prices and terms, customer data if it queries a CRM, and off-script answers under your brand.
How long does the initial scan of an assistant take?+
Working days, without touching your systems: only what is visible from outside. For remediation we need the minimum access, agreed in writing.
Is it GDPR compatible?+
Yes: the report points out what personal data the assistant exposes and how to fix it.
Start by knowing where you stand.
Tell us your situation and we will book an assessment. No commitment, with a clear plan at the end.
Book an assessment →