Safe AI use training for companies: leadership, technical teams and staff

Your team already uses ChatGPT, Copilot or Gemini. We teach them to get the most out of it without giving data away: three tracks, hands-on labs on 0data —the private AI platform we build and use every day— and real cases with their figures, not slides.

Last updated: October 2026

Scope your project · Design my company's training

Rather talk it through? Call us at +34 960 730 013

3 tracks — Leadership, technical teams and staff, each with its own level and format
0data — Labs on our own private AI platform, not on brochure demos
0 data — Leaves your premises during the labs: synthetic data or your own installation
Real cases — Published work, with its figures and methodology in plain sight

AI is already in your company. The judgment to use it isn't yet

AI assistants have crept into daily work without going through any approval process. They summarize contracts in minutes, write code, answer customers and sort hundreds of applications. The power is real, and that is why people use them.

So is the risk, and it is almost never in the tool but in the judgment: a client contract pasted into a personal account, a password left in a chat history, a résumé scored by a machine, a chatbot tricked into revealing instructions by a manipulated email, or a made-up figure that reaches the board without anyone asking where it came from.

Generic AI training changes none of that. What works is practicing with the business's own texts, watching the risks happen in a controlled environment and leaving with rules you can apply on Monday. That is why the labs run on 0data, a real platform with the personal data barrier, the audit trail and the local models working.

Would you paste it? Try it before the training

It is the first exercise in the staff session. Pick a work text, decide whether you would paste it into an AI assistant and see what happens on each path: a personal account on a public assistant, or 0data on your own server.

Client contract

Summarize this contract with Fictitious Turia Workshops Ltd (tax ID B12345678): 15% penalty if the service is down for more than 4 hours and automatic renewal on March 1.

On a personal account of a public assistant, no: it is client data under confidentiality. On an approved corporate tool or on 0data, yes. Telling them apart is the first thing you learn.

Public assistant, personal account:

The text leaves the company for a provider with no data processing agreement in place.
Depending on the account and its settings, the conversation is retained and may be used to improve the service.
A confidentiality commitment to the client has been broken without anyone noticing.

0data on your server: The model runs on your server and receives codes instead of the client and its tax ID. The summary comes back with the names restored and the query is logged in the audit trail.

Code with a password

Why does this connection fail? db = connect(host='10.0.4.12', user='admin', password='Summer2026!')

No, not even on an approved tool: a password pasted into a chat is no longer secret. If you already did it, change it today.

Public assistant, personal account:

The password and the internal address stay in the history of an account the company doesn't control.
If that account is compromised, the attacker gets the door and the key at once.
The right move now is to change that password, not to delete the conversation.

0data on your server: The barrier detects the credential and the IP before the model and swaps them for codes: the model explains the connection error without ever seeing the password.

A candidate's résumé

Score this candidate from 1 to 10: Marta Fictitious, 41 years old, two children, ID 00000000T, twelve years in logistics.

No. It is personal data from someone who hasn't consented, and using AI to select staff is a high-risk use under the EU AI Act. AI can help organize requirements; the decision about a person, no.

Public assistant, personal account:

A candidate's personal data sent to a third party with no legal basis.
Age and family situation can bias the score: automated discrimination.
Selecting staff with AI requires human oversight and documentation under the EU AI Act.

0data on your server: The model only sees codes instead of the name, age, family and ID: it compares experience with the job requirements and the decision stays with a person.

A customer complaint

Write a friendly reply to this complaint from Lucía Martín, ID 48291736K, about the charge on account ES91 2100 0418 4502 0005 1332. Phone 612 345 678.

Not like that. The reply needs neither the ID nor the IBAN: remove them and the AI helps just as well. It is the golden rule of the session: paste the problem, not the person.

Public assistant, personal account:

An ID number and an IBAN together are identity-theft material if that conversation leaks.
A customer's bank details at a third party with no contract: a GDPR problem, not an IT one.
The reply would have been just as good without any personal data.

0data on your server: Name, ID, IBAN and phone are replaced with codes before the model and the real values are stored encrypted elsewhere. The reply comes back written and with the customer's name in place.

Board minutes

Summarize the minutes: the board approves selling the north plant to a fund for €38M before the November 15 announcement.

No. It is a deal that isn't public yet. If it leaks, the problem is no longer IT: it is a market problem and a board liability.

Public assistant, personal account:

An unannounced deal leaves the company inside a text nobody controls.
In a listed company it is inside information: disclosing it has legal consequences.
Responsibility lies with whoever pasted it, not with the tool.

0data on your server: The summary is made on your machine, which can run with no outside connection. The asset, amount and date reach the model as codes and the query is sealed in the audit trail.

Illustrative examples: names, companies and numbers are fictitious.

What the training covers

What information never goes into an AI assistant, and how to use it just as well without it
Approved tools vs personal accounts: retention, training and who reads the conversations
Shadow AI: spotting the tools the team already uses without IT knowing
Prompt injection and over-privileged agents: how an AI application is attacked and defended
Pseudonymizing personal data before any model, local or external
Sourced answers: spotting made-up figures and demanding the origin of every data point
GDPR, EU AI Act and NIS2 obligations applied to daily use
What to do in the first minutes if sensitive information has already been pasted

How we prepare the training

Quick assessment: Which AI tools each team uses today, for what and with which data, in one call and a short questionnaire.
Tailored track and examples: We choose the tracks and build the exercises around the kind of documents your business handles.
Training with a lab: Sessions and workshops with labs on 0data: with synthetic data in our lab or on your own installation.
Follow-up: A report for leadership on what came up in the sessions and, if it fits, a six-week 0data pilot on your use case.

Three tracks, one judgment

Each audience needs something different: whoever decides, whoever builds and whoever uses AI every day. Pick one to see its modules and which labs run on 0data.

Leadership and boards

Board, senior management and department heads: whoever decides which AI is used and answers for it. · 2-3 hour session, on-site or online

What AI can and can't do today: A live, hype-free demo: where it really saves hours and where it is wrong with total confidence. Lab on 0data: We watch 0data answer with the source in front of it, and what it does when it has none: it doesn't assert.
Risk and liability: Data leaks, automated decisions and who is accountable: the EU AI Act, GDPR and NIS2 in board language.
Reading an AI-made figure: A figure without a control group misleads. We see it with our analysis of 17,746 insider purchases reported to the SEC: the real effect was half of the apparent one. Lab on 0data: We walk through that analysis step by step, with the public data and 0data's methodology.
Buy, build or wait: When a commercial tool is enough, when data can't leave the building and what to get in writing before signing.

Judgment to approve or stop an AI project, the questions to ask a vendor, and GDPR, EU AI Act and NIS2 obligations translated into decisions.

Technical teams

Development, systems, data and security: whoever integrates, deploys or codes with AI. · 1 or 2-day hands-on workshop, with lab

Attacking an assistant: prompt injection: How a chatbot is tricked into revealing instructions or data through an email, a PDF or a manipulated web page, and how to test it before going live. Lab on 0data: We attack and defend a 0data agent in an isolated environment, with booby-trapped documents.
Personal data before the model: Detecting and pseudonymizing names, ID numbers, IBANs or phone numbers before a text reaches a model, local or external. Lab on 0data: We configure 0data's barrier —what is blocked, what is flagged and what is pseudonymized— and test it with texts from your sector.
AI-generated code: The typical flaws in code an assistant suggests, dependencies that don't exist and secrets in the history, and how to review it without slowing the team down.
AI in production: a dashboard that doesn't lie: A process can finish successfully hundreds of times in a row without bringing in a single record. Monitor what it brings, not just whether it finishes. Lab on 0data: We reproduce the real case of a 0data source that spent 67 days empty with its process green, and build the alarm that catches it within hours.
Local models and sourced answers: Choosing between a local or commercial model, routing by cost and risk, and checking every figure against its origin before showing it. Lab on 0data: We query 0data over a test corpus and force the cases where it must answer that it has no source.

Knowing how to attack and defend an AI application, set up a personal data barrier and monitor an AI system in production so it doesn't fail silently.

All staff

Anyone who uses ChatGPT, Copilot or Gemini in their daily work. · 60-90 minute session, online or on-site, in groups

Would you paste it?: The game on this page, with texts from your own business: contracts, emails, code and résumés. Lab on 0data: We watch live as 0data swaps personal data for codes before the model reads it.
Getting value without giving data away: Summarizing, drafting, reviewing and translating with prompts that work, removing first whatever identifies people and clients.
Approved tool or personal account: What changes between the business and the free version: conversation retention, use for training and who can read them.
When AI gets it wrong, or is used to fool you: Made-up answers given with total confidence, and fake voices or videos asking for an urgent transfer. How to check.
If you already pasted it: What to do in the first minutes: delete, report and, if it was a password, change it.

A clear rule on what gets pasted and what doesn't, prompts that really save time and what to do in the first minutes if something was already pasted that shouldn't have been.

What the company receives

Sessions and workshops per track, online or on-site
Access to the 0data lab during the training
One-page guide: which data goes in which tool
Attendance record and evaluation per participant
Report for leadership: tools in use, frequent questions and risks detected
Onboarding material for new hires
Recording of the sessions, where agreed

What you gain, what you risk and how it's controlled

Six uses of AI already present in companies. Switch lenses to see each one from all three sides: the training covers all three, not just the fear.

Summarizing contracts and documents (risk: Medium). Power: What used to take a morning, in minutes: clauses, deadlines and penalties in plain view. Risk: Client data leaving the building and summaries that skip exactly the clause that matters. Control: Approved tool or local model, personal data removed and a human reading of whatever will be signed.
Coding with assistants (risk: High). Power: More speed on repetitive tasks, tests and documentation. Risk: Code that looks right and isn't, dependencies that don't exist and secrets pasted into the chat. Control: Code review, secret and dependency scanning and rules on which repository can be pasted.
AI customer service (risk: High). Power: Instant answers, at any hour and in any language. Risk: A chatbot that promises what the company doesn't offer, or is tricked into revealing instructions by a manipulated message. Control: Limits on what it can say, injection testing before going live and hand-off to a person when needed.
Recruitment (risk: High). Power: Sorting hundreds of applications by objective requirements. Risk: Bias, personal data with no legal basis and a high-risk use under the EU AI Act. Control: Documented criteria, minimal data and the decision about each person always in human hands.
Market and regulatory monitoring (risk: Medium). Power: Thousands of sources watched and an alert the same day: regulation, competitors, leaks. Risk: Made-up figures nobody checks and alerts nobody reads. Control: Every figure checked against its source, as 0data does, and alerts only when something is out of the ordinary.
Agents that act on their own (risk: High). Power: Chained tasks with no intervention: research, fill in, send. Risk: An over-privileged agent that obeys whatever a manipulated email or web page tells it. Control: Least privilege, human approval for anything with consequences and a log of every action.

When this training fits

Companies where the team already uses ChatGPT, Copilot or Gemini with no written rule
Committees that must approve or stop an AI project and want their own judgment
Technical teams about to integrate AI into a product, a chatbot or an agent
Organizations that already have an AI usage policy and need it applied
Companies that have already had a scare: sensitive data pasted into a public chat
Companies considering a private AI platform that want to see it working before a pilot

Real cases built with 0data, and what they teach

Published work, with its figures and methodology in plain sight. These are the cases we use in the training: what you learn from a real system sticks; what you learn from a manual doesn't.

The green dashboard that brought no data

67 days without a single document, with the process green every three hours

What was happening: A market source on our own platform stopped bringing in documents. Its process kept finishing successfully: 536 runs in a row, all green.
What 0data did: Since then 0data monitors every source against its own rhythm: if it goes quiet when it normally speaks, it alerts. With that rule, the alert fires within six hours.
What the training takes from it: Technical track: an AI system isn't monitored by whether it finishes, but by whether it brings what it should.

Anatomy of 0data

When an executive buys their own shares

17,746 insider purchases reported to the SEC, measured against a control group

What was happening: After an executive buys shares in their own company, the stock goes up. It looks like a crystal-clear signal.
What 0data did: We measured the same thing with a control group of the same stocks on random dates. Net of that control, the real effect is roughly half.
What the training takes from it: Leadership track: before deciding on an AI-made figure, ask "compared with what?".

When an executive buys their own shares

How a news story travels

124,802 headlines from 162 outlets, over fourteen days

What was happening: How long does a story take to move from one outlet to another? There is a lot of opinion and very little measurement.
What 0data did: Automated collection, headline clustering and measurement: some pairs of outlets publish in the same minute 84.8% of the time, against 1.9% for the rest.
What the training takes from it: Technical track: the hard part isn't the model, it's collecting, cleaning and monitoring the data it lives on.

How a news story travels

Intelligence briefs: one piece every weekday

5 weekly series generated from what 0data collects every day

What was happening: Actively exploited vulnerabilities, ransomware leak sites, CERT advisories and phishing sites: too much to read by hand every day.
What 0data did: Every weekday a short piece is generated from that data, with the figures computed and the source at the bottom.
What the training takes from it: Staff and leadership tracks: useful AI isn't the one that writes nicely, it's the one that puts the source under every figure.

Threat landscape, every Monday

Every Monday ransomware, exploited vulnerabilities, phishing and advisories, against a normal week

What was happening: Knowing whether this week is worse than normal requires comparing, not just counting.
What 0data did: 0data counts what happened and compares it with a normal week; what can't be asserted is flagged instead of made up.
What the training takes from it: All tracks: reading a threat figure without alarmism and knowing what to do with it.

Cases from our own platform and analyses published on the blog. Client projects are not published: in the training they are only used anonymized and with permission.

A generic AI course vs this training

What changes when the labs run on a real platform and the examples come from your business.

	Generic AI course	QuantumSec training on 0data
Examples	Textbook, the same for everyone	From your business and from real published cases
Labs	With public tools and sample data	On 0data: synthetic data or your own installation
Risks	One slide at the end	The thread of the whole session: leaks, injection, bias and made-up figures
Who delivers it	Generalist trainers	The team that audits AI systems and builds 0data
Afterwards	A certificate	Guide, onboarding material, leadership report and, if it fits, a 0data pilot

Compare with

Frequently asked questions about safe AI use training

Do I need to buy 0data to take the training?

No. The training is booked on its own and works just as well if you use ChatGPT, Copilot or Gemini. 0data is the lab environment because it lets you see the risks and the controls really working. If it fits afterwards, the way to start is a six-week pilot at a fixed price.

Do the labs use our data?

They don't need to. By default we work with synthetic data in our 0data lab. If you already have 0data installed or want to try it in-house, the labs run on your installation and the data never leaves your infrastructure.

Is this the same as the safe AI use service?

No. The safe AI use service is the full governance project: discovering which tools the team actually uses (shadow AI), drafting the usage policy and training on it. This page is just the training, by track, and can be booked without the rest.

How much does it cost?

A fixed price per session or workshop, depending on the track, the number of attendees and whether it is on-site or online. We give it to you after a first scoping call, which is free.

Do we need to already have an AI usage policy?

It is not required. If one exists, the training builds on it; if not, we give the basic rules — what never gets pasted, which tools are approved — without waiting for the document to be drafted.

How is it different from the training for technical teams or for leadership?

Those programs cover cybersecurity in general: secure development and defense, or board risk and liability. This one focuses only on AI: how to use it, how to attack it and how to control it, with labs on a real platform.

Related resources

Sources

Ready to protect your business?

Answer the scoping form at your own pace and we'll use it to prepare your proposal. If you'd rather talk first, book a free 30-minute call.

Scope your project · Design my company's training