AI-Powered Cyber Threats: What Changed and What Didn't
There's a lot of noise about AI-powered attacks, and it's worth separating the real change from the marketing. AI didn't invent new categories of attack. What it did was make the existing ones dramatically cheaper to run and much harder to spot — which turns out to matter a great deal for small businesses specifically.
The actual change: economics
Convincing, targeted attacks used to require human effort. Researching a target, writing a believable email in fluent English, referencing the right people and projects — that took an attacker an hour or two per target. At that cost, only large organizations were worth the trouble. Small businesses got the generic, badly-written spam that everyone had been trained to recognize.
That cost has collapsed. The research and the writing now take seconds. Which means the quality of attack that was once reserved for enterprises is now aimed at ten-person companies, because it costs essentially nothing to do so.
You are not more targeted because you're more interesting. You're more targeted because targeting got cheap.
What this looks like in practice
Phishing that passes every old test
The advice everyone learned — watch for bad spelling, odd grammar, generic greetings — described attackers writing in a second language. That constraint is gone. Messages now arrive fluent, correctly formatted, addressed to you by name, referencing your actual manager and your actual vendor, drawn from your website and LinkedIn.
The practical consequence is significant: "does this look suspicious?" is no longer a reliable test. Defenses have to be process-based rather than vigilance-based. We go through the specifics in phishing trends.
Voice cloning
This is the one that surprises people. A short sample of someone speaking — a conference talk, a podcast appearance, a voicemail greeting, a company video — is enough to generate convincing synthetic speech in their voice.
The attack: your bookkeeper gets a call that sounds like the owner, urgently authorizing a payment. Or an employee gets a call that sounds like IT, asking them to approve an MFA prompt. The voice is right, the urgency is plausible, and voice has always been the fallback people use to verify things.
The defense is a code word. Agree a phrase for verifying unusual financial requests by phone, known to a small group and never sent by email. It sounds low-tech because it is, and it works precisely because it's out of band.
Business email compromise at scale
Attackers who compromise a mailbox can now have a model read months of correspondence and produce a request that matches the writing style, references real ongoing work, and arrives at a plausible moment. The invoice fraud version — a real vendor, a real outstanding invoice, changed bank details — is the most expensive attack category facing small businesses.
Faster exploitation of known vulnerabilities
Turning a published security advisory into working exploit code has gotten faster. That shortens the window between a patch being released and it being exploited — which makes patching promptly more valuable, not less.
What hasn't changed
This is the reassuring part, and it's genuinely reassuring.
AI improved the delivery of attacks. It did not change what an attacker has to do once inside: authenticate as someone, run code on a machine, move between systems, reach your data. Every one of those steps is still blocked by the same controls as before.
- MFA still stops credential theft. An AI-written phishing email that harvests a password still hits an MFA prompt.
- Conditional access still stops session theft. A stolen cookie replayed from an unmanaged device still gets rejected.
- EDR still catches behavior. AI-assisted malware still has to encrypt files or steal credentials, and those actions look the same.
- Patching still closes the door. Faster exploitation of a vulnerability you've already patched doesn't help the attacker.
- Immutable backups still enable recovery. Ransomware doesn't care how well-written the email that delivered it was.
There is no new product you need to buy to counter AI attacks. The fundamentals hold — what changes is that human judgment is no longer a reliable layer, so the technical and process controls have to carry more weight.
What to actually change
- Move verification off email entirely. Payment changes and unusual requests get verified by voice on a known number, or by an agreed code word. Written channels can be impersonated; a callback to a number you already had cannot.
- Stop training people to spot bad writing. Retrain toward "verify unusual requests regardless of how legitimate they look," and toward reporting fast without embarrassment.
- Enable number matching on MFA so there's nothing to blindly approve.
- Deploy conditional access for at least administrators and finance staff.
- Consider security keys for anyone who can move money. They're phishing-resistant by design, and no amount of AI polish gets around cryptographic site verification.
On the defensive side
It's worth noting the same technology is being applied to defense, and effectively. Behavioral detection, anomaly spotting across large volumes of sign-in data, and automated response are all genuinely better than they were. Microsoft Defender and similar platforms benefit from this without you needing to do anything.
We build AI systems ourselves — we've written about putting agents inside our own accounting system — and the honest summary is that the defensive gains roughly track the offensive ones. The businesses that struggle aren't the ones facing smarter attacks; they're the ones that never got the fundamentals in place.
Want to know if your defenses hold up?
Vulcan365 provides managed IT and security for businesses across Birmingham and Central Alabama. We'll walk through which of these attacks your current setup would actually stop — and which it wouldn't.
Book a security review