Cover image: AI Security: What to Audit Before Scaling
IA

AI Security: What to Audit Before Scaling

Five fronts, a 30-day roadmap, and the myths that are costly. What to audit before scaling the use of AI in the company.

Por Agência Kaizen6 min read

Compartilhar

AI security is the set of controls that prevents a model from delivering sensitive data, inventing information as if it were fact, or being manipulated by malicious text. The minimum audit has five fronts: what data enters, who sees what comes out, what tools the model can trigger, how each response is recorded, and who reviews before publishing. A company that skips this step scales an invisible risk — and it's cheap to resolve at the beginning.

The question has shifted from "are we going to use AI?" to "with what limits?". By 2026, almost every commercial operation will have an assistant writing emails, analyzing WhatsApp conversations, or filling out CRM. The risk is not that the model is dumb; it is that it is obedient to an instruction pasted by someone who shouldn't.

The five real risks

  1. Data leakage. Someone pastes a contract, a CPF, or the margin spreadsheet into a public assistant. The information leaves your control and does not return.
  2. Instruction injection. A text coming from outside — page, email, customer message — contains commands that the model follows, ignoring the system's rules.
  3. Confident hallucination. The model responds with invented numbers and sources, in the same tone as the truth. In public content, this destroys authority.
  4. Uncontrolled automation. The agent sends a message, creates a discount, or deletes a record without human approval.
  5. Nonexistent trace. There is no record of who requested what, what the model responded, and what was published.

Minimum audit, block by block

Front Control question Expected evidence
Datawhat categories enter the prompt?list with personal and financial data marked
Accesswho can use which tool?profile with minimum permission
Outputwhat does the user see of the reasoning?screen without exposed internal data
Actionwhat can the AI execute on its own?list of what requires human approval
Recordwhat is recorded and for how long?log with date, author, and content

1. Classify the data before writing the prompt

Separate what can go to an external model from what cannot. Customer personal data, contract information, margin, and strategic planning belong in the second group. From this list, you decide where each assistant runs: public version, corporate version with retention off, or internal model.

2. Treat third-party text as hostile input

Web page, attachment, comment, and WhatsApp message are untrusted input. The model should receive this content as data, never as instruction. In practice: a separate block in the prompt, an explicit rule that there are no valid commands inside, and human review when the response involves publishing, sending, or altering records.

3. Human approval for all external actions

Simple rule that resolves most issues: AI can research, draft, and suggest; publishing, sending, scheduling, connecting, and altering sensitive data requires approval. It’s the difference between an assistant and an intern with the bank password.

4. Record each relevant interaction

Log with date, user, summarized prompt, response, and what was done afterward. Without a record, there is no way to investigate incidents or prove compliance when a client asks.

5. Combine the legal basis, do not improvise

Anyone handling customer personal data needs to know the legal basis, retention period, and response channel to the data subject. Generative AI does not create exemption: data remains data — see data privacy in marketing.

Do you use AI in customer service and don’t know what is being recorded? Get a free diagnosis with Agência Kaizen.

A 30-day roadmap

  • Week 1: inventory. List all assistants in use, even those someone opened on their own.
  • Week 2: classification. Mark what each one receives and generates; block what it shouldn't.
  • Week 3: rules. Human approval for external action, interaction recording, incident channel.
  • Week 4: testing. Simulate an injection case and a leakage case. If the response comes out without barriers, control does not exist.

Example from the segment

A customer service operation with automated WhatsApp was receiving messages from customers with links to a page. The page contained hidden instructions asking the assistant to reveal the sales script and offer a discount. Since the external text was treated as instruction, the agent complied. The correction was in design: external content began to enter in an isolated block, and any discount began to require human approval. The model did not change; the process changed.

How to measure maturity without complicating

Score from 0 to 2 for each front — data, access, output, action, and record. A sum below 6 indicates an uncontrolled operation; between 6 and 8 requires a plan with a deadline; above 9 is ready to grow. Repeat every quarter, as each new tool changes the design.

The model is not the only risk

Half of the incidents that arise in practice do not originate from the model, but from the extension: a plugin that reads the entire email box, an agent connected to the CRM with administrator permission, integration that stores the access key in plain text in the code.

Before auditing the prompt, audit the connection:

Connection Risk Minimum control
Email boxreading third-party messages as instructionreading permission restricted to folder
CRMaltering or exporting the entire databasestep-by-step access, not administrator
Key and tokenleakage in log or repositorysecret in vault, quarterly rotation
Storagedocument copy without deadlinedefined retention and automatic deletion

This framework is the tedious part and the one that prevents headaches the most: each permission that the agent does not need is an accident that does not happen.

Myths that cost dearly

  • "We only use the corporate version, so we are safe." The corporate version does not prevent someone from pasting data they shouldn't.
  • "The model is good, it doesn't make mistakes." Every model makes mistakes and does so with conviction.
  • "Recording is bureaucracy." Recording is what allows responding to the client and the audit on the same day.
  • "Security is only the IT department's job." Those who operate daily define the data that enters.

FAQ

Do I need a written AI usage policy?

Yes, and it should be brief. One page outlining what is allowed and what is not, where human approval is mandatory, and who to report incidents to is worth more than a 40-page manual that no one reads.

Can I use ChatGPT or Gemini with customer data?

Only if the tool's contract guarantees no use for training and defines retention, and if the data falls within the category allowed by your policy. When in doubt, anonymize.

How do I prevent the agent from making mistakes on its own?

Minimum permission and approval for external actions. Start with the agent suggesting and the human clicking.

Is AI security expensive?

The design takes weeks, not vendor licensing. The expensive part is the incident: exposed contract, undue discount, post published with customer data. Also see AI governance and AI in customer service.

Start where it hurts the most

Start with the data. If you don't know what information is currently entering your assistant, that is the only item on the list to resolve this week. Then, ensure human approval for all external actions. Together, these two steps address most of the risk that is concerning and can be done without changing tools.

Also read:

The Kaizen Agency implements AI agents with data control, human approval, and interaction logging. Talk to us for a diagnosis of what is currently running.

Also read: 6 ITEMS THAT MUST NOT BE MISSING.

Quer Aprender Mais?

Junte-se à Universidade Kaizen e tenha acesso a cursos gratuitos, trilhas de aprendizado completas e conteúdo exclusivo para profissionais que querem dominar as melhores práticas do mercado.