Skip to content
← Insights

The AI Breach Wave Is Coming. Most of It Will Be Self-Inflicted.

November 11, 2026 · 6 min read

AI Summary
Generating summary…

My prediction for 2027: more breaches than any year on record, with AI sitting in the middle of most of them.

That part's easy. Everyone is saying it. The part nobody in a vendor booth wants to say out loud is that most of those breaches won't come from some genius AI attacker. They'll come from us. From companies that bolted AI onto everything in eighteen months and skipped the boring work that keeps data where it belongs.

The numbers already back this up.

The trend line isn't subtle

IBM's 2026 Cost of a Data Breach Report found that one in four malicious breaches is now AI-enabled, up 56% in a single year. Those breaches cost an average of $6 million, roughly a million more than the global average of $4.99 million. Mostly deepfake impersonation and AI-built malware.

That's the attacker side. Now look at ours.

Shadow AI showed up in 43% of incidents in IBM's 2026 research, up from 20% the year before. It more than doubled in twelve months. And 92% of organizations hit by AI breaches lacked basic access controls.

Ninety-two percent.

That's not a sophisticated adversary. That's an unlocked door with a welcome mat.

The attacker got faster. We didn't.

Cogent Research looked at more than 69,000 CVEs and found attackers needed about 125 days to build a working exploit in January 2025. By April 2026, with AI doing the heavy lifting, that dropped to half a day. Synack's numbers tell the same story: mean time from CVE to first exploitation went from 56 days in 2024 to under 10 hours.

Now think about how most organizations still run. Monthly patch cycle. Annual pentest. A vulnerability scan that gets reviewed when someone has time. Verizon's 2026 DBIR says vulnerability exploitation was the way in for 31% of breaches.

Your patch window was designed for an attacker who needed two months. That attacker retired.

We're writing the vulnerabilities ourselves now

Vibe coding is the other half of this. Describe an app, let the AI build it, ship it when the tests go green.

One benchmark of AI coding agents on real-world tasks found that more than 70% of functionally correct agent-generated code contained exploitable vulnerabilities. Georgia Tech's Vibe Security Radar tied 56 CVEs to AI-generated code in the first quarter of 2026 alone.

Full disclosure: I build with these tools. I've shipped products this year that would have taken a dev team six months. The speed is real. But when the AI hands me an authentication flow, I know what I'm looking at. Most people shipping vibe-coded apps don't.

Every small business with a nephew and a Cursor subscription is now a software vendor. Nobody's reviewing the code. "The tests pass" quietly became the security review.

Your AI assistant is the new insider threat

Then there are the agents. We gave them read access to SharePoint, the CRM, the shared drive and the CFO's inbox, and connected them to the internet.

Earlier this year a security startup pointed an autonomous agent at McKinsey's internal AI platform. It got in within two hours using SQL injection, a bug class that's been documented since the 1990s, chained with a broken access control. Microsoft and Salesforce both had to patch prompt injection flaws in Copilot Studio and Agentforce that let attackers pull data out through public-facing web forms.

Think about what we actually did. We hired the most gullible employee imaginable, one that follows instructions from any document, email or calendar invite it reads. Then we gave it more access than most of our staff.

I've worked more than 600 incident-response engagements over 26 years. The root cause is almost never exotic. Over-permissioned accounts. Stale credentials. Something exposed that nobody knew about. AI doesn't change that pattern. It multiplies it. An over-permissioned human can only read so fast.

Three opinions that'll annoy somebody

"Shadow AI" is a cop-out. It isn't shadow when leadership set AI adoption targets and never gave anyone a policy. That's sanctioned negligence with a softer name. If your board approved a Copilot rollout before anyone reviewed who can access what, your board approved the breach. They just haven't seen the invoice yet.

A chunk of the AI security market is selling the cure for a disease the same industry is spreading. AI-powered defense works, and I'll get to that. But you can't buy a product that undoes giving an agent domain admin. Vendors lead with the dashboard because the dashboard sells faster than "go fix your permissions."

Cyber insurance is about to get ugly. Insurers priced a lot of current policies against exploit timelines that no longer exist. My bet: by 2027 renewals you'll see AI-specific exclusions and questionnaires asking exactly what your agents can touch. If you can't answer, expect the premium to answer for you.

The honest caveat

Not every data point says doom. VulnCheck's first-half 2026 report suggests AI is, for now, more of an advantage for defenders when it comes to finding and fixing bugs. IBM found that organizations using AI and automation in their security operations cut breach costs by nearly $2 million.

I believe both of those. AI rewards the organizations that already did the fundamentals and punishes the ones that didn't. The gap between those two groups is about to get very wide, very fast.

What to do on Monday

Not a 40-point framework. Just the stuff that moves the needle:

  • Find every AI tool touching company data, including the ones nobody approved. You can't govern what you can't see.
  • Treat every agent like a new hire with zero judgment. Its own identity, least privilege, everything logged.
  • Assume anything an agent reads is hostile input. Email, PDFs, calendar invites, web forms, all of it.
  • Put someone with real security chops between AI-written code and production.
  • Shrink the patch window on internet-facing systems to days, and test continuously instead of once a year.
  • Tabletop an AI agent compromise with your executives before you have to live one.

Bottom line

AI isn't going to break into your company. You're going to let it in, hand it a badge, and forget to check which doors it opens.

The breach wave is coming. Whether you're in it is still mostly your call.

Frequently asked questions

Will AI cause more data breaches?

Yes. IBM's 2026 Cost of a Data Breach Report found one in four malicious breaches is now AI-enabled, a 56% jump in one year. The bigger driver is organizations adopting AI tools and agents faster than they secure them, with shadow AI showing up in 43% of incidents.

What is shadow AI?

Shadow AI is any AI tool employees use with company data without formal approval or security oversight. It's dangerous because nobody is tracking what data goes into those tools or what access they have.

How fast are attackers exploiting vulnerabilities with AI?

Research from Cogent and Synack shows exploit timelines have collapsed from weeks or months to hours. Monthly patch cycles and annual pentests can't keep up with that pace.

How can businesses reduce AI-related breach risk?

Inventory every AI tool touching your data, give AI agents least-privilege access with their own identities, treat everything an agent reads as untrusted input, require security review of AI-generated code, and shorten patch windows on internet-facing systems.

Next Step

Ready to strengthen your organization's resilience?

A 30-minute discovery call to discuss your cybersecurity posture, incident readiness, and whether advisory support is the right fit.