AI now makes SMBs worth a hacker’s time

COMMENTARY: For years, small- and midsize businesses (SMBs) had an unspoken advantage in the social engineering game: they weren’t worth the effort.

Crafting a convincing phishing campaign took time, research, and at least passable English. A sophisticated attacker wasn’t going to spend three hours personalizing a spear-phishing email to trick a controller at a 200-person manufacturing company when a large Fortune 500 target was sitting right there. That math no longer works.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

There’s a reason the Nigerian prince email became a punchline rather than a full-blown crisis. For nearly two decades, phishing largely relied on volume over quality, and human beings turned out to be surprisingly good at sniffing out the fakes.

Awkward phrasing, bizarre formatting, a sender claiming to be the “Microsoft Security Department” using a Gmail address—our built-in BS detectors caught most of it before we ever needed a security awareness program to tell us what to look for. Even people who couldn’t articulate why an email felt wrong could usually feel it was wrong.

Generative AI broke that defense. Large language models let threat actors generate context-aware, grammatically flawless lures in any language, at any scale, for any target. Phishing attacks increased 1,265% through late 2024, and the messages today don’t look like the Prince of Lagos asking to help him move $40 million out of his country.

Much like email and text-based phishing, voice phishing has undergone the same transformation. Vishing attacks surged 442% over the same period, driven by AI voice-cloning tools that can reproduce a person’s speech from as little as seven seconds of publicly available audio.

And even though frontier models like ChatGPT, Claude, and Google Gemini come with a mountain of guardrails that can make it difficult to use AI to clone a voice, local models can do this kind of cloning without making a laptop break a sweat. An attacker doesn’t need a recording studio or a voice actor anymore. They only need a YouTube clip of the company’s CEO speaking at a chamber of commerce event and an open-source voice-cloning tool on a laptop.

The resulting call sounds close enough to authorize a wire transfer, change direct-deposit details, or convince a Help Desk representative to reset a password over the phone. For organizations that still rely on “I recognized their voice” as an informal verification step, that defense has evaporated.  

Then there’s the automation piece that gets far less press coverage right now, but may matter more in practice. AI-driven reconnaissance and scanning tools have compressed what in the past required weeks of manual attacker research into hours. Targets get fingerprinted, attack surfaces get mapped, and vulnerable services get probed at a pace that beats most SMB security teams’ ability to notice, let alone respond.

The same AI that writes the phishing email can identify which employee to target, determine which system they’re likely to have access to and craft the pretext that will get them to click. It’s not a single tool to do one task well. It’s an automated pipeline doing everything adequately, and adequately is more than enough when the target isn’t expecting it.

The shift from automation to attack chaining

That automated pipeline has become more capable. A little over a year ago, the human was still doing most of the work: asking a model for an Nmap scan or a piece of code, going back and forth with it, and eventually copying and pasting something useful. By late 2025, AI was being used more as an architect or planner. The attacker could give it a target and ask for an attack plan, and the model could work through a series of tasks to try. The human was still involved in initiating the intent and giving instructions, but AI offered more of the scaffolding around the attack.

The next step has been more consequential because the individual stages of an attack aren’t really what changed. Reconnaissance, initial access, execution, privilege escalation, lateral movement, exfiltration, and impact have all been around for years, and most individual steps can already get automated. Today, attackers have the ability to chain those actions together, end-to-end and at machine speed. Instead of needing explicit instructions at every stage, the model can reason from one step to another toward a goal and make more of those decisions autonomously.

That significantly shortens the cycle. AI can make thousands of decisions per minute where a team of human attackers might spend weeks working through different attack chains. The economics change again when the attacker isn’t simply automating reconnaissance or generating a better phishing email, but reducing the human effort between the stages of an attack.

We’re already seeing examples of what that can look like. In an April presentation, Anthropic’s head of threat intelligence described the company’s analysis of nearly 800 malicious actors it had identified using Claude over the previous year. In a separate case disclosed by Anthropic, a threat actor used Claude Code to automate significant portions of an extortion campaign, including conducting reconnaissance, stealing data, generating ransom notes, and reasoning about how much money to demand. If the victim was a church, for example, the model might conclude that $75,000 made more sense than $2.5 million.

The uncomfortable reality for SMBs today: AI has eliminated the economics that once protected them. When an attacker can target 1,000 organizations for the same cost it once took to target one, every business with a registered domain name and an email server has now become worth the effort.

Scale makes ordinary vulnerabilities more dangerous

The most sophisticated AI capabilities will make the headlines. We’re already seeing models capable of finding novel vulnerabilities that traditional pattern-matching tools miss, reasoning across codebases, and chaining weak signals into working exploits. But that doesn’t mean an automated attacker will start with the hardest possible way into an environment.

It will first ruthlessly scan for the low-hanging fruit. Missed software patches. Exposed network settings. Simple misconfigurations. Unrotated credentials. Forgotten subdomains. Those are the things that actually hurt teams, and none of them are new. It’s the boring stuff. The difference here is that AI can look for them faster, across far more organizations, and without nearly as much human effort.

That’s why I wouldn’t respond to AI-enabled attacks by throwing out everything we already know about security. The blocking and tackling matters more when an attacker can cheaply and continuously look for the tasks the team missed. Understand what the external attack surface looks like. Keep up with patching and credential rotation. Make sure one compromised system can’t become a path through everything else.

And assume some attacks will get through. Prevention still matters, but so does controlling the blast radius when prevention fails. That’s a larger conversation, particularly as autonomous AI begins operating inside organizations as well as attacking them from the outside.

For now, we need to focus on a much simpler point: AI doesn’t need to invent an entirely new class of cyberattack to create a much bigger problem. For SMBs that could once take some comfort in not being worth the trouble, that math no longer works.

Jim Sherlock, vice president of AI and cybersecurity R&D, ProCircular

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.