The right mental model
AI hasn't invented entirely new attack categories. It has industrialised and personalised the attacks that were already working. The phishing email that tricked someone in 2024 still tricks people in 2026 — it's just better written, more targeted, available in more languages, and produced at one-tenth the cost.
The defensive implication: most of what protected you in 2024 still works in 2026, but it has to be applied more rigorously. Two-factor authentication everywhere becomes non-negotiable. Awareness training shifts from “spot the typos” to “follow the verification procedure regardless of how legitimate the request looks.”
The six threats
AI-generated phishing at scale
Generative AI lets attackers produce flawless, context-aware phishing emails in dozens of languages, at zero marginal cost. Volume of phishing campaigns is up materially year-over-year; quality of individual messages is past the point where 'look for typos' is useful advice. Defence: advanced email security with behavioural detection, MFA everywhere, awareness training that focuses on verification procedures rather than visual cues. See our deep-dive on AI Phishing.
Voice and video deepfake BEC
Attackers now clone executive voices from a few seconds of audio (LinkedIn videos, podcast appearances, conference recordings) and make voice calls authorising fraudulent wires. Video deepfakes for live Zoom impersonations are catching up. Defence: out-of-band verification procedures for any wire transfer or payroll change, regardless of who appears to be requesting. See Deepfake BEC.
Shadow AI data leakage
Employees paste client data, contracts, and confidential information into consumer AI chatbots (ChatGPT, Claude.ai, Gemini, Copilot) to get help drafting emails or summarising documents. That data may train future models, sit in chat history, or surface in another user's session. Defence: written acceptable-use policy, blocklist or sanctioned alternative for AI tools, awareness training, M365 Copilot or Gemini for Business deployment for legitimate need. See Shadow AI.
AI-augmented credential stuffing
Attackers use AI to analyse breach data, generate plausible password variants, and target specific people more effectively than brute force ever could. Defence: phishing-resistant MFA (FIDO2 / passkeys), unique passwords via a password manager, dark-web credential monitoring, conditional-access policies.
Prompt injection against business AI
Where businesses are deploying AI into customer-facing workflows (chatbots, support agents, internal copilots), attackers craft malicious prompts that override the AI's instructions and exfiltrate data, generate fraudulent content, or invoke unsafe tools. Defence: input sanitisation, output filtering, principle of least privilege for AI agents, monitoring of AI interactions for anomalies.
Synthetic identity fraud
AI generates convincing fake identities — names, photos, employment histories, social media presence — that pass casual verification. Used in account-opening fraud, fake-vendor onboarding, and CV fraud during hiring. Defence: stronger vendor onboarding (background checks, video verification), payment-process verification, monitoring of customer-onboarding patterns.
What changes in your security programme this year
- Awareness training rewrite. “Look for typos and bad grammar” is dead advice. Replace with: “Verify any payment, credential, or sensitive-data request out-of-band, regardless of how it looks.”
- Written AI acceptable-use policy. If you don't have one, your employees are using consumer AI tools with company data right now. We cover this in AI Governance Policy for SMBs.
- Phishing-resistant MFA for privileged accounts. AI-assisted credential phishing defeats SMS and push-based MFA. FIDO2 security keys or passkeys for admin accounts.
- Out-of-band verification baked into financial workflows. Wire authorisations, vendor payment-method changes, payroll direct-deposit changes — all require a callback to a known number, not just an email or call response.
- Sanctioned AI tools. If employees have legitimate AI needs, give them an enterprise tool (M365 Copilot, Gemini for Business) instead of forcing them into consumer alternatives. See M365 Copilot Security for SMBs.
What hasn't changed
Many of the controls that mattered in 2024 still matter in 2026 and always will:
- MFA on every account.
- Modern endpoint detection & response on every device.
- Immutable offsite backups, tested for recovery.
- Email security with attachment sandboxing and lookalike-domain detection.
- Patch management on a documented timeline.
- Documented incident response plan, tested annually.
AI changes the threat distribution; it doesn't change the foundations. A small business with strong foundational controls and a written AI acceptable-use policy is materially better positioned than one chasing AI-specific tooling without the basics.
On the doom narrative
AI-ready security assessment
Free 30-minute consultation. We map your current controls against both the 2026 AI threat landscape and the foundational baseline, and tell you honestly where the gaps are.
Book the free assessment