What AI Has Actually Changed About Cybersecurity — And What It Hasn't

Every security vendor now has an AI story. Most of them are trying to sell you something, so it is worth separating what the incident data actually shows from what makes a good pitch deck.

The short version: AI has not invented many new categories of attack. What it has done is make the existing ones cheaper, faster, and far more convincing — which turns out to matter more than a new attack type would have.

The economics changed, not the playbook

Business email compromise, credential theft, and pretexting were the top attack patterns before generative AI, and they still are. Verizon’s 2025 Data Breach Investigations Report — covering more than 22,000 incidents and 12,195 confirmed breaches — found that 60% of breaches involved a human element, with credential abuse (22%) and vulnerability exploitation (20%) as the leading initial access vectors. Pretexting made up 27% of social engineering breaches, nearly all of it business email compromise.

None of that is new. What changed is the cost of running those attacks well.

The old advice — watch for bad grammar and awkward phrasing — worked because convincing English at scale used to be expensive. It is not anymore. Industry reporting now puts more than 80% of phishing emails as containing AI-generated content, and CrowdStrike’s threat research recorded an 89% increase in attacks from AI-enabled adversaries during 2025.

A phishing email that references your actual vendor, your actual invoice cycle, and your CFO’s actual writing style is not caught by proofreading. That single shift invalidates a decade of security awareness training built around spotting typos.

Voice and video are now part of the attack surface

The genuinely new development is impersonation of people, not just companies.

IBM’s research on AI-enabled attacks found deepfake impersonation was the most common type at 45%, ahead of AI-enabled malware (19%) and phishing campaigns (17%). Google Cloud’s 2026 Cybersecurity Forecast expects attackers to accelerate “vishing (voice phishing) with AI-driven voice cloning to create hyperrealistic impersonations of executives or IT staff.”

Note the second half of that sentence. The impersonation target is often IT staff, not executives — because the fastest route into most organizations is convincing a help desk to reset a password or enroll a new MFA device. A cloned voice defeats “I recognize them on the phone” as a verification step.

This is the control most businesses have not updated. If your password reset process depends on someone sounding right, it is already broken.

Identity is where the breaches happen

If you only fix one thing, fix this. Google Cloud cites Cloud Threat Horizons data showing over 70% of cloud breaches stem from compromised identities, and forecasts that IAM failures will be “the primary initial access vector for significant enterprise compromise.”

AI makes this worse in two directions at once. Attackers get better at harvesting and using credentials, while organizations quietly add a new population of accounts — AI agents and integrations — that often sit outside normal identity governance entirely.

IBM’s data makes the gap concrete: 13% of organizations reported an AI-related security incident, and 97% of those acknowledged they lacked proper AI access controls.

The risk that arrives without a purchase order

Whatever your AI policy says, your staff are already using AI tools. Google Cloud’s forecast is blunt about it: “Whether your company adopts agents or not, your employees will use them for work,” and notes that “merely blocking them simply won’t work (they will sneak back in immediately).”

The practical exposure is not exotic. It is a staff member pasting client data into a consumer chatbot, or connecting an unsanctioned tool to your Microsoft 365 tenant with broad permissions. For firms with HIPAA, PCI, or contractual confidentiality obligations, that is a compliance event, not just an IT concern.

What this means if you are not a bank

Most published threat research is written for enterprises with security teams. The translation for a 20-to-200 person business is narrower than it looks, because attackers are not choosing between you and a bank — they are running the same automated campaigns against everyone, and you are cheaper to hit.

Verizon’s SMB findings are the uncomfortable part: ransomware lands disproportionately on small and mid-sized businesses, with a median ransom payment of roughly $115,000. That is a business-ending number for a lot of companies, and it does not require anyone to have targeted you specifically.

Five controls do most of the work:

1. Phishing-resistant MFA. Not SMS. App-based push with number matching at minimum; hardware keys or passkeys for administrators. This addresses the credential abuse that leads the breach statistics.
2. A verification step your help desk cannot be talked out of. Password resets and MFA re-enrollment need an out-of-band check that does not rely on recognizing a voice. Write it down, make it non-negotiable, and test it.
3. Least privilege that includes non-human accounts. Service accounts, API keys, and AI integrations need owners, scopes, and expiry dates like anyone else.
4. Backups you have actually restored from. Tamper-resistant, off-network, and tested — because prevention that assumes nothing gets through is not a plan. This is the control that decides whether an incident is an outage or a closure.
5. Awareness training that reflects current attacks. Stop teaching typo-spotting. Teach verification behavior: how to confirm a payment change, what a legitimate IT request looks like, and who to call.

Where AI genuinely helps defenders

The same technology is doing real work on the defensive side, and this is not vendor theater. Google Cloud describes security operations shifting “from a monitoring hub into an engine for automated action,” with AI handling alert grouping, correlation, and incident summarization so analysts can spend their time on decisions rather than triage.

For a smaller organization, that capability arrives through your provider rather than through a platform you buy and staff. It is a fair question to ask any managed IT partner: what is actually monitoring your environment at 2am, how quickly does a real detection reach a human, and what happens next.

The honest summary

AI has not made security unwinnable. It has removed the cost advantage defenders quietly relied on — that running a convincing attack at scale took skill and effort.

The fundamentals still hold. Identity, patching, backup, and verified process were the right answers before, and they remain the right answers now. What has changed is that doing them approximately is no longer good enough, and the window between exposure and exploitation keeps shrinking.

Reviewing your security posture against these threats? Get in touch and we will walk through where your current controls stand — or estimate what an outage would actually cost you with our downtime calculator.

Sources: Verizon 2025 Data Breach Investigations Report · Google Cloud 2026 Cybersecurity Forecast · Microsoft Digital Defense Report 2025 · CrowdStrike Global Threat Report