Attackers hired agents. So should you.

For most of its history, an attack took a person. Someone had to research the target, write the message, make the call, and wait. That human cost was a quiet form of protection, because people are expensive and there are only so many of them. Most of us were safe simply because we were not worth the hours. That protection is now dissolving, and it is worth understanding exactly how, because the fix is calmer than the headline.

In late 2025, Anthropic reported a state-linked espionage campaign against roughly thirty organizations in which an AI system performed an estimated 80 to 90% of the operational work, with humans stepping in only at a few decision points. Some researchers questioned the details, and that skepticism is healthy. But the direction is not seriously in doubt. An earlier case had already used an AI agent to run an extortion scheme against seventeen organizations. The labor that used to limit attacks is being automated away.

Why automation changes the math

When the work becomes nearly free, three things change at once. Attacks get cheaper, so more people are worth targeting, including ordinary households that were never worth a human's time. Attacks get more patient, because an agent can wait weeks, reading and learning, without losing interest. And attacks get more personal, because the same system that writes a generic lure can just as easily write one tuned to you, your employer, and the people you trust.

That last point is measurable, and it is the number that should get your attention.

Automated lures get clicked far more often

Share of targets who clicked, in testing reported by Microsoft.

In plain terms: in Microsoft's testing, a standard phishing message was clicked by 12% of targets, while an AI-automated one was clicked by 54%, roughly four and a half times as often. Source: The Register, citing the Microsoft Digital Defense Report 2025, October 2025.

The same cheapness applies to voices and faces. A report cataloguing verified cases counted 1,567 deepfake incidents and $1.28 billion in losses in 2025, and that counts only public, confirmed events, so it is a floor. The most expensive single example is already famous: the engineering firm Arup lost about $25.6 million after an employee joined a video call in which every other participant was a convincing fake of a senior colleague. The FBI has separately warned of AI-generated voice messages impersonating senior US officials. A face on a screen and a voice on a line are no longer proof of anything.

What does not change

Here is the part the fear tends to drown out. Automation makes attacks cheaper and more numerous, but it does not invent a new way into your life. The agent still has to get a human to click, to pay, or to tell it something. The pressure still arrives as a message, a call, or a login page. And the defenses that work against a patient human work just as well against a patient machine, because they do not depend on you spotting the fake. They depend on you refusing to act on the fake's word alone.

You will not win by detecting the forgery. You win by building steps that a forgery cannot complete on its own.

Keep that sentence in mind, because it is the whole strategy. Every durable defense below is a step the attacker cannot finish without you, no matter how good the imitation is.

What an individual or a family can actually do

Decide that voices and faces are not identification. This is the mental shift that does the most work. A caller who sounds like your son, or a video of your boss, proves nothing by itself now. Agree in your family and your business that identity is confirmed by a shared secret or a call back on a known number, never by recognition alone. A family passphrase is the simplest version, and a cloned voice cannot supply it.

Require two people for money that moves. The most effective single rule against automated fraud is also the oldest in accounting: any payment above a threshold you set needs a second person to approve it, through a separate channel. The Arup loss happened in a single call. A rule that the transfer waits for a colleague's independent confirmation defeats the deepfake even when the video is flawless, because the agent cannot be in two places or two channels at once.

Protect the master accounts and assume patience. Because agents will wait and watch, the accounts that reset everything else, your email and your phone carrier above all, deserve your strongest protection: a passkey or a hardware key, not a text code that can be redirected. Assume an attacker who gets a foothold will sit quietly for weeks. The goal is to give them nothing worth waiting for.

Slow the moment down. Automated attacks still lean on urgency, because urgency is what stops verification. A patient machine will happily manufacture a crisis to make you hurry. The counter is unchanged and unglamorous: when money or secrecy or speed appears, pause, and verify through a channel you chose rather than the one that contacted you.

Why "so should you" is not a slogan

If the attacker's advantage is a tireless system that tests your defenses constantly, then the honest answer is a tireless system on your side that tests them first. This is the reasoning behind how we are built. Every Feint member, at every tier, has an agent fleet that runs every night against their own accounts and devices, looking for the gap before anyone else does, with a person reviewing anything it finds.

The agents handle the part that benefits from never sleeping: the constant, boring, exhaustive checking. The people handle the part that still needs judgment: deciding what matters, and running the monthly exercise that tells a family whether its own rules actually hold under pressure. Neither half is enough alone, which is precisely the lesson the attackers have taught us. In one controlled test, an AI phishing agent outperformed an elite human red team by 24%, and in the real campaigns the machines did the volume while humans made the calls. The defense has to be shaped the same way.

None of this requires you to match the attacker's tools. It requires you to adopt the one thing their automation cannot get around: steps that need a real, deliberate human choice, made slowly, that a confident fake can never make for you. The machines got faster. The way through is still to refuse to be rushed, and to build a few checks that only you can complete.

What to look out for

In an age of automated attackers

1. A voice or video as proof of identity. Confirm by a shared secret or a call back on a known number, never by recognition.
2. A payment with a single approver. Require a second person, on a separate channel, for any money above a set threshold.
3. Urgency engineered to skip a step. A manufactured crisis is a tell; a real one survives a short pause to verify.
4. Text-message codes on your most important accounts. Prefer passkeys or hardware keys, which cannot be redirected.
5. Quiet, patient contact over weeks. Automated attackers wait; treat a slow, friendly build-up with the same care as a sudden demand.
6. Defenses you have never tested. Knowing your rules is not the same as knowing they hold; rehearse them before the real attempt.

Sources

1. The Hacker News, on an AI-orchestrated espionage campaign, 2025-11; and Anthropic, detecting and countering misuse, 2025-08-27.
2. The Register, citing the Microsoft Digital Defense Report 2025, 2025-10-16.
3. Resemble AI, 2025 Deepfake Threat Report, 2026-03.
4. Business Insider, on Arup's $25.6 million deepfake loss, 2024-05.
5. NBC Chicago, on the FBI warning about AI voice impersonation, 2025-05.
6. Hoxhunt, AI phishing versus human red teams, 2025-03.