Introduction
The 2026 Microsoft Digital Defense Report describes a threat environment where familiar attack methods are becoming faster and easier to scale with AI.[2] Microsoft highlights identity compromise, interconnected software, cloud infrastructure and AI systems themselves as important attack surfaces.[2]
The useful lesson for an SMB is not that every attacker has suddenly become a supercomputer. It is that weak credentials, excessive permissions and unmanaged tools can now be discovered and exploited more efficiently. Basic security controls still matter, but the order of operations matters too. Identity is the front door to email, cloud files, ERP, banking workflows and AI agents.
What happened
Microsoft published its report findings on 13 October, based on observations from July 2025 through June 2026.[2] The company says attackers are increasingly using AI to identify vulnerabilities, generate malware and conduct phishing and online fraud at scale. It also warns that the data and infrastructure supporting AI systems are becoming targets.[2]
This is a vendor report based substantially on Microsoft’s telemetry and analysis. It should be read as Microsoft’s view of activity visible across its ecosystem, not as a complete census of all cyber incidents. The regional page includes figures about Indonesia, but those figures should not be generalised to Belgian or European SMBs.
The broader operational recommendations are still relevant: use phishing-resistant authentication, apply least privilege, improve visibility over AI applications and agents, reduce exposure on internet-facing systems, test recovery and prepare for incidents.[2]
Why it matters to business
Most SMB attacks do not begin with cinematic hacking. They begin with a reused password, a fraudulent sign-in, an over-permissioned account, an exposed remote service or a convincing message sent to someone under pressure.
AI adds two complications. First, fraudulent messages can be produced quickly in the language and style of a target organisation. Second, businesses are connecting AI tools to real systems. An agent that can read a shared mailbox, update CRM records or prepare payments has an identity, permissions and credentials, whether the organisation manages it that way or not.
That changes the security question. It is no longer enough to ask whether employees have multifactor authentication. Leaders must ask which non-human identities exist, what each one can reach, where its secrets are stored and which actions require human approval.
Business continuity is part of the same problem. If email, files or ERP access is disrupted, can the company still invoice, pay staff, serve customers and recover cleanly? A backup that has never been restored is a comforting theory, not a control.
Practical actions for the next 30 days
- Protect privileged accounts first. Require phishing-resistant MFA or passkeys for administrators, finance users and remote access. Remove dormant accounts.
- Inventory human and agent identities. Include service accounts, API keys, bots, integrations and AI agents. Name an owner for each.
- Reduce permissions. Give every user and agent only the access required for its task. Separate read, draft, approve and execute permissions.
- Control high-impact actions. Payments, supplier changes, bulk exports and destructive operations should require explicit approval or a second channel.
- Find shadow AI. Review browser, identity and expense data for unapproved tools. Offer an approved route instead of relying on prohibition alone.
- Patch exposed systems. Prioritise internet-facing services, remote access, identity infrastructure and software with known exploitation.
- Test one recovery scenario. Restore critical data into an isolated environment and measure the time needed to resume a core process.
- Run a payment-fraud drill. Simulate a supplier bank-account change or executive request. Verify the callback and escalation process.
A practical containment pattern
Treat every operational agent as a junior system user. Give it a unique identity, a narrowly defined role and short-lived credentials where possible. Log its inputs, tool calls and outputs. Allow autonomous execution only for low-impact, reversible actions. Route unusual amounts, new recipients, bulk changes and sensitive records to a person.
Monitoring should focus on behaviour, not only login failures. Flag impossible travel, unusual export volume, access outside normal hours, repeated denied actions and sudden changes in tool use. When an alert fires, the response team should be able to disable the identity without taking the entire business offline.
How VDS can help
VDS Consultoria designs AI-powered operational systems with controls built into the workflow. We can map identities and integrations, reduce agent permissions, add approval gates, centralise audit trails and connect alerts to a tested response process.
We do not sell cybersecurity as a standalone service. We make the AI automation and digital products we build safer to operate, easier to inspect and simpler to contain when something behaves unexpectedly. Security becomes part of the system design, not an appendix added before launch.
Sources
[2] https://news.microsoft.com/source/asia/2026/10/13/microsoft-digital-defense-2026-report-highlights-new-cybersecurity-challenges-amid-rapid-ai-adoption — Microsoft Digital Defense 2026 Report Highlights New Cybersecurity Challenges