How to reduce your risk of becoming a victim

By Ivo Donker · 7 November 2026 — written with AI assistance (Claude Opus 5.5); facts checked against the sources listed below

Most AI attacks don’t involve any wizardry, but rather old vulnerabilities at high speed: a reused password, a router that hasn’t received an update for two years, someone paying under time pressure. As a result, solid basic security still works, and your own AI can even help you. Below are three levels: home, the AI homelab and the SaaS company. Start with the level that suits you and carry forward the steps from the previous level.

Level 1 — at home and in the family

Agree on a code word

Obtaining a voice sample of your child, partner or manager only takes a fraudster a few seconds of audio from social media or a voicemail. The best defence is old‑fashioned: agree with your close ones on a code word or verification question that isn’t posted online. If “your daughter” calls in panic from a new number asking for money, hang up and call her back on the number you already have. The same applies at work: always verify a payment request received by phone, WhatsApp or video call via a second, known channel.

Passkeys and a password manager

AI can generate phishing emails that are flawless and personalised; you can no longer spot them by spelling mistakes. Passkeys (logging in with your fingerprint, face or device PIN) are resistant to phishing because they only work on the genuine website. Enable them for your email, bank, Apple/Google/Microsoft accounts and social media. Where that isn’t possible yet, use a password manager with a unique password for each site, plus two‑step verification via an app rather than SMS.

Automatic updates, retire old devices

Attackers scan the entire internet and keep track of which device is running which software version. Turn on automatic updates on your phone, computer, router, NAS and smart devices. A camera or router for which the manufacturer no longer releases updates should not be connected to the internet. Check your router on the manufacturer’s website and disable “remote management” and UPnP if you don’t need them.

Be careful with what you allow your AI assistant to do

Don’t paste passwords, citizen service numbers or medical data into a chatbot. Don’t give an AI assistant that reads your mail or calendar the authority to make payments or send messages on its own. A malicious email can contain hidden instructions for your assistant; this is called prompt injection.

How AI can help you properly

Ask your AI assistant to assess a suspicious email, SMS or link before you click. Paste the text, not your login credentials.
Have it create a step‑by‑step plan for the security settings of a particular device or account.
Use the check at Have I Been Pwned and ask your AI what to do after a password has been leaked.
If you’re unsure about a payment request or a scam, submit it to the Fraudehelpdesk (Fraud Helpdesk) (088-724 73 72).

Level 2 — the AI homelab

A homelab with local models, a NAS, Docker containers and agents that run code is more interesting to an attacker than you might think: plenty of computing power, API keys with credit, and often a tunnel to the outside. The key steps:

Know your attack surface. Make a list of everything reachable from outside: ports, tunnels, subdomains, dashboards. Look at it the way an attacker would: look up your own IP address and domains on Shodan or run the free scan from Internet.nl over your domain and email. Close anything you haven't deliberately opened.
Nothing manageable directly exposed to the internet. NAS management, Docker dashboards, Ollama, n8n or Home Assistant ports belong behind a VPN (for example WireGuard or Tailscale) or a tunnel with mandatory login (such as Cloudflare Access). An open Ollama port is found within a day.
Agents in a cage. Let a coding agent or AI assistant work in a separate user, container or VM, with only the folder it needs. No access to the Docker socket, your entire home folder or your NAS shares. Use flags such as "skip permissions" only in a disposable environment.
Short-lived keys. Give each agent its own API key with a spending limit, and revoke it when the agent is done. Don't put keys in prompts, repos or agent memory; use a secrets manager or your operating system's keychain.
Check what you install. An MCP server, plug-in or npm package is code running with your rights. Check who maintains it, pin versions and be extra alert to package names an AI recommends you but that turn out not to exist (slopsquatting).
A backup the attacker can't reach. At least one copy offline or immutable (snapshots with a lock period, a disk that stays disconnected). Test the restore every quarter.
Follow the vulnerabilities that are actually being exploited. Subscribe to the advisories from the NCSC (National Cyber Security Centre) and check weekly whether anything on the KEV list is running in your environment. That list is short enough to keep track of and exactly what attackers work with.

How to set up local models safely is also in the guide to local LLMs.

Level 3 — SaaS companies and development teams

For companies that provide software, the requirements have been mandatory since 2026. The Cyber Resilience Act obliges manufacturers to report actively exploited vulnerabilities within 24 hours, and the Cybersecurity Act (NIS2) imposes care and reporting duties on a large part of the business community. You can only achieve a 24‑hour report if you know which components run where.

Level 3 — SaaS companies and development teams
Measure	Why now	Minimum
Inventory of assets and dependencies	Exploits are once again the number‑one entry route	SBOM per release, owner per system, list of external exposure
Phishing‑resistant identity	Identity is the door for both people and agents	Passkeys/FIDO2 for all administration, SSO, no shared admin accounts
Patch timelines	Time to exploitation is hours, not weeks	Critical KEV vulnerabilities within 48 hours, high severity within 7 days
Agent and MCP policy	OWASP Agentic Top 10: goal hijack, tool misuse, overly broad privileges	Unique identity per agent, allow‑list of tools, approval required for writes, payments and deletes, kill‑switch
Supply‑chain integrity	Third parties play a role in almost half of the breaches	Lockfiles, signed builds, provenance checks, supplier assessment
Detection and logging	Attackers remain undetected for weeks on average	Centralised, immutable logs; a rehearsed incident playbook
Recovery	Ransomware targets backups first	Separate management path, immutable backups, measured recovery time

Also employ AI for defence: have a model summarise and deduplicate alerts, check pull requests for secrets and unsafe patterns, and attack your own application in a test environment with prompt‑injection scenarios. Ensure that the decision to block, roll back or pay is always made by a human.

The background to this table and the deadlines up to 2030 are in the look‑ahead to 2030.

If it goes wrong

Money sent to a scammer: call your bank immediately (often a payment can still be reversed), file a report with the Politie (Police) and notify the Fraudehelpdesk (Fraud Helpdesk).
Account compromised: change the password using a different, clean device, log out of all sessions and check forwarding rules in your email.
Business incident: businesses can turn to the NCSC (National Cyber Security Centre) (which has incorporated the Digital Trust Center (Digital Trust Centre) since 2026); personal data breaches must be reported within 72 hours to the Autoriteit Persoonsgegevens (Authority for Personal Data).