SMS from the Tax Authority: How Neural Networks Made Phishing Indistinguishable from Truth

You surely remember those times when a phishing message could be recognized from a mile away by ridiculous typos, strange syntax, and the general feeling that the text had been translated from an exotic dialect through an early version of Google Translate. Those times are officially in the past. Today, with tax season in full swing, scammers are armed with large language models, and this has turned ordinary spam into a high-precision weapon of social engineering. Now an SMS about a supposed tax deduction owed to you looks as if it had been checked by an entire department of copywriters and lawyers. And this is far more frightening than the mere prospect of a tax audit.

The problem isn't that scammers have become smarter, but that the tools have become more accessible. Previously, conducting a large-scale campaign required people with good language knowledge and understanding of bureaucratic style. Now, a single prompt to a hypothetical "dark" GPT is enough to generate thousands of variations of convincing texts that perfectly imitate the tone of official authorities. These models don't just correct errors; they adapt context to a specific country, region, and even current news, making the trap as relevant as possible for the victim. When you receive a notification about funds being returned on the very day you expected it, resisting the urge to click the link becomes almost impossible.

Why is this important right now? We are at a tipping point where classic security advice like "check the text for errors" is becoming not just useless, but dangerous. It creates a false sense of security. If the text is written perfectly, it doesn't mean it's authentic. It means the scammer has access to a modern neural network API. The FTC and other regulators continue to issue advisories, but they are catastrophically falling behind the pace of generative AI development. Scam schemes are now scaling at the speed of light, and what was a single case yesterday becomes an epidemic affecting millions of users in a matter of hours.

The connection is direct: the development of LLMs has lowered the barrier to entry for cybercrime to a historical minimum. You no longer need to be a hacker or linguist; you just need to know how to use a chatbot. This creates a colossal burden on the filtering systems of telecom operators and email services, which are forced to implement AI to combat AI-generated content. We're observing a kind of "arms race" in the text space, with your smartphone becoming the battlefield. And while defensive algorithms learn to distinguish the synthetic politeness of a neural network from genuine officialdom, the only reliable shield remains your personal skepticism.

In the near future, we will see even more sophisticated attacks. Imagine a neural network analyzing your public social media data and sending you an SMS from the tax authority, taking into account your actual line of work or recent major purchases. This is no longer science fiction, but a matter of setting up the right script. Technology allows phishing to be personalized at an industrial scale, and this changes the rules of the game for everyone who uses the internet. We will have to get used to the idea that any incoming message demanding quick action with money is a fake by default, no matter how convincingly it looks.

Key takeaway: You can no longer trust your eyes — modern LLMs write official letters better than live officials. If a message asks you to click a link to receive money, close it and log into your account on the official agency website manually. Are you ready for your next "conversation" with your bank or tax authority to be completely generated by an algorithm?