
Cybersecurity and Superintelligence for Companies: What to Protect First
Superintelligence is still hypothetical, but AI agents already reach credentials, channels and customer data. Here is the order to protect each asset.
Want to apply this in your company?
Book a call with Juan →Superintelligence, in the sense of systems that would outperform humans at most tasks, does not exist yet, and there is no consensus on when or whether it will. For your company's security, the useful question is different: what can an AI agent reach today inside your systems? The practical answer is to protect credentials first, then permissions, irreversible actions, outbound channels, databases and, last, the records that let you investigate any incident.
Key takeaways
- An agent is only as risky as the credentials and permissions it receives, so protection starts with the secrets vault and the minimum scope of each integration.
- Irreversible actions, such as charges, cancellations, mass sends and changes to customer records, need human approval before they run.
- The WhatsApp number and the email domain are reputation channels and also entry points for fraud, not just ways to contact people.
- Databases with personal data should be accessed through limited queries, never through full access, which helps with LGPD compliance and reduces the damage of a mistake.
- Without a record of decisions and actions, no incident can be reconstructed, so logging belongs on the protection list from the first pilot.
Why start with what the agent can reach
The risk of an AI agent rarely comes from an intelligence running out of control. It comes from three factors combined: the agent has access to real systems, it processes text that you do not control, and it can take actions on behalf of the company. A customer message, an attachment or a web page can contain instructions written to mislead the model, a technique known as prompt injection. OWASP, which maintains a list of risks for applications built on language models, covers this problem and also excessive agency, when a system receives more power than the task requires.
For that reason, agent security starts with an inventory of what the agent can touch, not with the choice of model. This is the most useful question for a manager: if the agent fails or is manipulated, what is the worst possible outcome with the permissions it has today?
Practical protection order: from credential to database
The order below follows a simple rule: protect first whatever gives access to everything else, then whatever causes immediate and irreversible damage, and last whatever stores accumulated information. Each item should be reviewed before the agent goes into real operation.
- Credentials and API keys: stored in a secrets vault, never in the prompt, example code or shared spreadsheets, with a defined rotation schedule.
- Identity and permissions: each agent has its own identity, with minimum scope per tool and per operation, separate from people's accounts.
- Irreversible actions: charges, cancellations, deletions, refunds and mass sends require a recorded human confirmation.
- Outbound channels: WhatsApp number, email domain and social media accounts, with volume limits and rules about who can trigger messages.
- Databases with personal data: queries by purpose, only the fields needed, read access separated from write access.
- Records and logs: every conversation, tool call and decision with date, origin and result, kept for a defined period.
- Configuration and instructions: prompts, rules and flows under version control, with a clear list of who can change them and how changes are reviewed.
The list does not cover everything, but it works as a starting point for a conversation with your technology team and with whoever is accountable for the process.
Credentials and permissions: how to limit the damage
A credential with broad access is the most valuable asset for anyone who wants to use the agent against the company. In practice, this means three precautions. First, the agent should use tokens with restricted permissions and an expiry date whenever the platform allows it, following the least privilege principle described in OWASP guidance. Second, each integration should be authorized separately, so that permission to read the calendar does not also bring permission to change customer records. Third, the company needs to know who can create, replace or revoke these credentials, and that answer should be written down.
A common mistake is giving the agent the same administrative account the team uses, for convenience during testing. That account can survive the pilot and become a permanent risk.
Irreversible actions and outbound channels
Some actions the agent should not execute on its own, even when it technically can. Refunds, price changes, contract cancellations, record deletions and sends to a large customer list are examples. For these, the right design is for the agent to prepare the action, show a summary to a responsible person, and execute only after a recorded approval.
The same care applies to channels. A WhatsApp Business number or an email domain that suffers abuse can be blocked, and the lost reputation takes time to recover. Before enabling outbound messages, confirm the consent base and the platform rules, as we cover in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act.
Databases and LGPD: access by purpose
The customer database is usually the most sensitive asset an agent can reach, because it combines names, phone numbers, addresses, purchase history and, in some sectors, health or financial data. The LGPD requires purpose, necessity and security in data processing, and the Brazilian National Data Protection Authority (ANPD) publishes the main guidance on the subject. For an agent, this translates into specific queries: an agent that reports the status of an order does not need the full purchase history, and an agent that schedules a visit does not need to read internal notes.
Care is also needed with decisions that significantly affect the customer, such as denying a service or flagging an order as fraud. These decisions require human review and an explanation, a topic covered in LGPD Automated Decision-Making: The Right to Review (Art. 20).
Records, testing and what to do when something goes off plan
Records are what turn an incident into learning. Every tool call should leave a line with the date, the request received, the action taken and who approved it, when applicable. Without this, the team cannot tell whether a problem was a model failure, a configuration error or a deliberate attempt at manipulation.
Illustrative scenario: a custom furniture store uses an agent on WhatsApp to report order progress. A customer pastes into the chat a text asking the agent to ignore its rules and send the list of clients from the week. If the agent only has access to the order status of an identified customer, the attempt finds nothing useful and is logged as a security event. If the agent has access to the entire CRM, the problem becomes a data leak. The difference lies in the permissions defined before the pilot, not in a well-written instruction inside the prompt.
Testing with manipulation scenarios, run in an isolated environment before real operation, helps find these gaps. The subject is covered in Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide.
How to apply this without turning security into bureaucracy
In our method, protection comes in four steps. First, map the assets the agent can reach, before any pilot. Then, limit data and actions to the minimum the chosen flow needs. Next, pilot a real flow with logs and a success metric agreed with the business. Finally, hand over the operation and decide whether the agent should be expanded, adjusted or stopped. This path does not eliminate risk, but it makes clear what was tested, who is responsible for each action, and what to do when something goes off plan.
The practical recommendation for the leadership team is simple: choose one flow, list what the agent can reach in it, cut whatever is not necessary, and require records of everything before widening the scope. To better understand what the word superintelligence actually covers today, read Superintelligence vs ChatGPT: The Real Difference for Business.
Frequently asked questions
Does superintelligence already represent a risk for my company?
Superintelligence, as a concept, does not exist yet, and there is no consensus on when it will. The concrete risk today comes from AI agents with access to real systems, credentials and data, which can be manipulated by customer text or carry out actions with excessive permissions. For that reason, protection should start with what these agents can already reach.
What is the first asset I should protect?
The first is the credential, because it gives access to all the other assets. Keep keys and tokens in a secrets vault, create a dedicated identity for each agent, and limit its scope to what the flow requires. Only then move on to action permissions, channels and databases.
Do I need to prevent the agent from taking any action on its own?
Not necessarily. The agent can handle low-risk, reversible tasks, such as checking a status or suggesting a reply. For irreversible or high-impact actions, such as charges, cancellations and mass sends, it is best that the agent prepares the action and a responsible person approves it before execution, with a record of that approval.
Where should my team start if it has no experience with agents yet?
Start with a single real flow that carries low risk and little sensitive data, and map what the agent can reach before switching it on. Define a success metric, keep logs from day one, and review permissions at the end of the pilot. If the result is good, widen the scope gradually; if it is not, adjust or stop without hesitation.