Every company that invests in information security starts from the same premise: the adversary is a person trying to trick a system. That premise is starting to be wrong. The phishing email landing in an inbox today may not have been written by anyone — it was generated by a language model that never misspells a word, is never in a hurry, and tests variations until it finds the one most likely to slip through unnoticed.
The problem isn’t just that attacks got more sophisticated. It’s that the defense logic most companies bought — comparing every file against a list of already-known threats — was designed for an adversary that no longer exists.
The attack also has a model behind it
A generative language model does in seconds what once required a malicious writer with time on their hands: it drafts a grammatically flawless phishing email, in the right tone for the victim’s industry, personalized with public information about the company — a real supplier’s name, the recipient’s job title, even the approximate value of a recurring invoice. The classic red flags any security training teaches employees to spot — bad grammar, a generic greeting, exaggerated urgency — simply stop existing, because the model learned to avoid them.
The same mechanism applies to malicious code: a generative model rewrites the same malicious function with a different syntactic structure every run — a different variable order, a different way of obfuscating the same logic — so the final file never has the same hash twice. That’s not a minor technical detail. It’s the piece that breaks the most common defense model on the market.
Picture the finance team of a mid-size company receiving an email from a recurring supplier, asking to update the bank details for the next payment. The writing is flawless, the invoice number cited is real, the tone matches every previous exchange. No antivirus flags anything, because there’s no malicious attachment and no link to a domain already known as a scam — it’s just a well-written email requesting a registration change. What stops the payment from going to the wrong account isn’t the list of known threats. It’s a system that learned that specific supplier’s normal behavior and flags that, in years of history, it had never once asked for a bank account change by email.
Signature-based defense already lost the race
A traditional antivirus works by comparison: every new file is checked against a database of already-catalogued threat signatures. That logic is efficient against reused attacks — and structurally blind to any attack generated in real time for a specific target, because that attack, by definition, has never been seen before. It’s not an occasional product flaw. It’s an architectural limitation: comparing against the past doesn’t protect against something designed to never resemble the past.
The structural alternative is behavior-based detection: instead of asking “is this file already known to be malicious?”, the system asks “does this behavior deviate from this user’s, this supplier’s, this account’s normal pattern?”. A model that learns an operation’s routine — which supplier charges what, how often, from which account — catches the deviation even without ever having seen that specific attack before. It’s the same shift in logic any AI system applied to decision-making requires: moving from “this has happened before” to “this is consistent with what should happen”.
It’s the same reading João Paulo Batistella, an innovation executive who has held roles at Ericsson, Telefônica/Vivo, and served as CEO of EISA for five years, makes when mapping the sectors AI should redesign completely in Brazil, not just make marginally more efficient: cybersecurity is on that list, because the threat stopped being “malicious person versus company” and became “attacker model versus defender model”. Measuring the two with the old ruler — a list of known signatures, updated once a month — no longer makes sense when the other side of the table also keeps learning.
3 questions a board should be asking about cybersecurity today
1. Does our defense recognize what it has already seen, or does it learn what normal behavior looks like for our own operation? If the answer is only the first, the company is protected against last year’s attack, not today’s.
2. When was the last time we tested our defense against an AI-generated attack, rather than a two-year-old checklist? An outdated security test doesn’t measure the real risk — it measures the risk the company no longer faces.
3. When an attack gets through — it’s not a hypothesis, it’s a matter of time —, how long does it take us to notice, and what’s the plan for after, not just the plan to try to prevent it beforehand? Perfect prevention no longer exists in this game; detection and response time is the metric that actually separates a contained incident from a crisis.
None of these three questions is technical in the sense of requiring code knowledge — they’re judgment questions, the kind a board should ask about any risk capable of threatening business continuity. Treating cybersecurity as solved because “the antivirus is installed” is the same trap as treating AI as an IT topic: delegating to the technical team a decision that is, at its root, about how much risk the company is willing to accept.
Information security stopped being a product you buy once — an annual license, an updated signature database — and became a capability that needs to keep learning, at the same pace as the attack it faces. A company that still measures its protection by its list of known threats is, in practice, unprotected — not because it did nothing, but because it did the wrong thing very well.
Follow Eleva Tecnologia for more analysis on technology applied to business: follow @ElevaTechno on X, @elevatechnologies on Instagram or Eleva Technologies on LinkedIn, or learn about the group at elevatec.net/about.
