How Basic Security Stops the Latest AI-Driven Attacks

AI has made attackers faster and cheaper, but it has not changed where they get in. The paths still run through identity, email, unpatched software, exposed systems, over-permissioned data, forgotten accounts, and a lack of visibility. If those are in good shape, most AI-assisted attacks have far less to work with. This article maps the newest AI-driven attacks to the foundational controls that stop or limit them, and shows how ThreeShield delivers each one with Lavawall®.

What AI actually changes

The evidence points to acceleration. The Canadian Centre for Cyber Security assesses that AI is lowering barriers to entry and improving the quality, scale, and precision of malicious activity, including personalized phishing at scale. Its June 2026 statement adds that frontier AI models can help threat actors find and exploit vulnerabilities, shortening the time defenders have to respond. The UK National Cyber Security Centre's 3 November 2026 assessment judges that AI-powered cyber tools have almost certainly led to a significant increase in the discovery of zero-day vulnerabilities, and that internal patching and testing processes will need to evolve to keep pace. Anthropic reported that attackers manipulated an AI coding tool to do most of an espionage campaign against roughly thirty targets, succeeding in a small number of cases, though the tool also hallucinated some credentials. And OpenAI's February 2026 report says threat actors typically use AI alongside more traditional tools.

The Verizon 2026 Data Breach Investigations Report found that 31% of breaches now start with software vulnerabilities, ahead of stolen passwords. Speed matters more than novelty, and speed is answered by fundamentals.

AI-accelerated attacks and the basics that answer them

1. File share, SharePoint, and Google Drive monitoring

Once inside, an account, human or automated, goes for data. Lavawall tracks opens, edits, downloads, shares, and deletions in SharePoint and OneDrive monitoring, and extends to on-premises file monitoring and Google Drive. Unusual downloads and risky external shares become visible, and so do the rapid, wide file changes typical of ransomware on a share. See also Google Workspace monitoring.

2. User approvals and access review

Least privilege limits what a stolen login, or an agent running as a user, can reach. Lavawall access reviews for SharePoint, Teams, and Entra ID route each entitlement to an approver who records keep, remove, or escalate, with a timestamped audit trail.

3. Conditional Access monitoring and controls

Microsoft recommends blocking device code flow wherever possible after tracking the Storm-2372 campaigns (Microsoft Threat Intelligence, updated 31 July 2026). Microsoft Learn advises auditing existing use first, allowing the flow only for well-documented, secured cases, and turning a new block policy on after a report-only run. Lavawall logs Conditional Access changes, and can maintain a policy that allows sign-ins only from the countries a client works from, report-only first. Read how we block device code flow. Details: Conditional Access monitoring and Microsoft 365 security monitoring.

4. User activity benchmarking and anomaly reporting

A stolen session often looks valid until you compare it with the person. Lavawall compares sign-in addresses and daily sign-in and file activity against each user's own baseline, and shows 7, 30, or 90-day trends. See user activity anomaly detection.

5. Strong MFA monitoring

MFA only helps if it stays strong. Lavawall alerts on new MFA methods in Microsoft 365 and flags weak or missing 2-step verification in Google Workspace. Microsoft recommends phishing-resistant methods such as FIDO tokens or Microsoft Authenticator with passkey, and Conditional Access authentication strengths can require them. Its 2026 Digital Defense Report again names phishing-resistant MFA and passkeys among the top safeguards. The Microsoft Digital Defense Report 2025 reports AI-automated phishing with click-through rates up to 4.5 times higher than traditional phishing. See MFA monitoring.

6. Outlook phishing reporter and analysis

Better-written phishing needs a faster response. The Phishing Reporter for Outlook add-in lets users report a message in one click and see an immediate risk read, with no extra sign-in or OAuth consent. Pair it with phishing simulations.

7. Unused account monitoring

Forgotten accounts are quiet, privileged footholds. Lavawall surfaces dormant and orphaned accounts, including inactive Google Workspace users and accounts that outlived a role change, so they can be disabled. See unused account monitoring.

8. Agentic AI monitoring

Unlike a chat website, an AI agent can open files and run commands. agentic AI discovery lists agent tools on managed Windows computers, the folders and connected tools each can reach, and set-ups that act without asking. It reads no prompts or files and blocks nothing. OWASP's 2026 Top 10 for LLM applications ranks Excessive Agency third. SaaS Discovery finds AI services used in the browser.

9. Patch monitoring

The Project Glasswing update says progress is now limited by how quickly flaws can be verified, disclosed, and patched, and advises network defenders to shorten their patch testing and deployment timelines while relying on baseline controls such as hardened defaults, MFA, and logging. Expect more frequent updates. The Canadian Centre for Cyber Security's frontier AI guidance recommends preparing for an increase in patching frequency, and Microsoft's Security Response Center expects releases to keep trending larger. Tenable counted 113 CVEs in Microsoft's January 2026 Patch Tuesday, 569 in July, 964 in September, and 1,407 in October, against 167 in October 2025 (other trackers differ by a few). The Windows team says customers will see a higher volume of security updates in each release as AI helps defenders find more issues. How much of the rise is due to AI is our inference, not something these counts show. Lavawall provides cross-platform patch management across Windows, macOS, and Linux.

10. Web vulnerability monitoring

Attackers, and their tools, look at your public site first. The domain security scan views your website, email, and DNS from outside, and the WordPress security scan flags outdated core, plugins, and themes. It is free to try.

The 10-point baseline checklist

1. Enforce MFA for every user, phishing-resistant where possible.
2. Block device code flow and legacy authentication in Conditional Access.
3. Review Conditional Access changes and MFA method changes.
4. Disable unused and orphaned accounts.
5. Run access reviews with named approvers.
6. Monitor SharePoint, OneDrive, Drive, and file-share activity.
7. Give staff a one-click way to report phishing.
8. Patch operating systems and applications on a short cycle.
9. Scan your websites and domains from outside.
10. Inventory the AI agents and tools in use, and limit what they can reach.

Frequently asked questions

Does AI create entirely new kinds of attacks?

Mostly it speeds up and cheapens familiar ones. OpenAI's February 2026 threat report says threat actors typically use AI alongside more traditional tools. The Canadian Centre for Cyber Security assesses that AI is lowering barriers to entry and improving the quality, scale, and precision of malicious activity.

Is multi-factor authentication enough against AI-written phishing?

Not by itself. Device code phishing, for example, can obtain tokens after a user completes a real MFA prompt. Phishing-resistant MFA, Conditional Access that blocks risky sign-in flows, and monitoring for MFA changes work together.

Why does AI mean we should patch more often?

AI tools are finding software flaws faster than maintainers can fix them. Anthropic's Project Glasswing update says progress is now limited by how quickly flaws can be verified, disclosed, and patched, and advises network defenders to shorten patch testing and deployment timelines. The Canadian Centre for Cyber Security recommends preparing for an increase in patching frequency, and Tenable counted 1,407 CVEs in Microsoft's October 2026 Patch Tuesday against 167 a year earlier.

What is agentic AI monitoring?

It means knowing which AI agents and coding assistants are installed on your computers, which folders and connected tools they can reach, and which set-ups act without asking first. Lavawall does this for managed Windows computers. It does not read prompts or files, and it does not block tools.

We are a small business. Where do we start?

Start with identity: strong MFA for everyone, a Conditional Access baseline, and removal of accounts nobody uses. Then confirm your patching is current, review who can reach your shared data, and give staff an easy way to report suspicious email.

Can any tool prevent every AI-driven attack?

No. These controls reduce the paths in and make misuse easier to see and contain. Treat them as a baseline that needs ongoing attention, not a guarantee.

Do you need Lavawall for this?

No. The controls are what matter. Lavawall is how ThreeShield delivers and monitors them for the organizations we co-manage, across many tenants from one console.

Last reviewed 2026-11-06. This article reflects public sources available on 6 November 2026. No control prevents every attack.