AI-Powered Cyberattacks

Four primary sources frame the shift, and none of them is a security vendor.

IBM, July 2026. IBM's 2026 Cost of a Data Breach study found one in four malicious breaches were AI-enabled, a 56% increase over the prior year, mostly deepfake impersonation and AI-enabled malware. Those breaches averaged about $6 million against a $4.99 million global average. Organizations using security AI and automation saved $1.93 million per breach.

Verizon, May 2026. The 2026 Data Breach Investigations Report found vulnerability exploitation was the initial access vector in 31% of breaches, overtaking stolen credentials for the first time in 19 years. Credentials still appear somewhere in 39% of breaches, and third parties were involved in 48%. Verizon says AI is shortening the window to exploit known vulnerabilities from months to hours.

Anthropic, November 2025. Anthropic reported disrupting a campaign it attributed with high confidence to a Chinese state-sponsored group. The operators used Claude Code as an agent against roughly thirty targets, with AI performing an estimated 80 to 90% of the campaign and humans stepping in at four to six decision points. A small number of intrusions succeeded. Three months earlier, Anthropic's August 2025 threat report described one criminal using an AI coding agent to extort at least 17 organizations, and another selling AI-built ransomware for $400 to $1,200.

Canadian Centre for Cyber Security, June 2026. The Cyber Centre's statement on frontier AI models warns that these models can shorten defenders' response windows "from days or weeks to hours."

Skill is no longer the bottleneck. Speed is.

This is the table we use in client briefings. The opinion column is where we part ways with a lot of vendor marketing.

Reconnaissance used to cost an attacker hours per target. Now a model can read your careers page, engineering blog and staff profiles and return a brief: cloud provider, VPN vendor, finance contacts, new hires who will not yet recognize the CFO's writing style. You cannot erase public data, but you can stop advertising your stack. Job ads listing exact firewall and VPN products are a gift.

When every lure is grammatically perfect, asking staff to spot red flags is a losing game. The control that holds is the one CISA calls the only widely available phishing-resistant option: FIDO/WebAuthn authentication. Our phishing scams guide covers the reporting habits that still matter.

For US federal agencies, CISA's BOD 22-01 sets a default of two weeks to fix newer entries on the Known Exploited Vulnerabilities list. Private companies are not bound by it, but we use it as a ceiling for anything internet-facing. If your VPN, firewall or file transfer appliance appears on that list, treat it as an incident clock, not a maintenance ticket.

Two details in the November 2025 report matter most. The AI hallucinated credentials and overstated its findings, so operators still had to verify its work. And every phase leaned on ordinary weaknesses: reachable systems, harvestable credentials, high-privilege accounts and data that could leave unnoticed. We walk through a comparable multi-day intrusion in anatomy of a five-day AI agent attack.

This is the section vendors leave out.

When you do not need an AI-specific security program. If you are a team of 15 on Google Workspace or Microsoft 365 with a handful of SaaS apps and no self-hosted servers, you do not need an AI threat platform. You need passkeys, auto-updates, managed endpoint protection and backups, and free or bundled tools cover most of it. A dedicated program makes sense once you run internet-facing infrastructure, hold regulated data, deploy agents of your own, or sell to enterprise buyers who will ask. If you deploy agents internally, start with how to secure AI agents.

Across our assessments, order matters more than budget. Here is the sequence we run.

Want a starting point? Our Cybersecurity Baseline Assessment measures where you stand against these controls.

In most companies we meet, the answer is "the CTO, in the evenings." That works until the exploitation window shrinks to hours. Someone has to own the patch clock, the alert queue and the tabletop. For many teams that is a fractional security leader rather than a $250,000-a-year full-time CISO. Our Virtual CISO service starts from $2,000 a month for 15 to 20 hours.

Yes, mostly as a side effect of scale. When reconnaissance and lure writing cost almost nothing, attackers choose targets by exposure, not size. An unpatched edge device at a 40-person company is as attractive as one at a bank.

Only after the basics. IBM's $1.93 million savings figure is real, but it measures security AI and automation inside a working security operation. Buying a standalone "AI defense" product before phishing-resistant MFA and fast patching is spending in the wrong order.

Not reliably yet. In the case Anthropic reported, humans still approved key decisions and the AI hallucinated results. Plan for attacks that are mostly automated, with a human checking the output.

Deepfake and voice-clone fraud target people and payment processes, not systems, and need their own controls. We cover them in The Dark Side of AI.

Not sure which stage your company is weakest at? Book a consultation and we will walk your attack chain with you.