AI cybersecurity for business now requires a closer look at identity, devices, and data access. Better-written phishing emails are only part of the challenge. Leaders also need to understand which systems an attacker could reach after compromising an account—and what company information employees are sharing with AI services.
Anthropic’s September 2026 threat report is attracting attention in today’s technology coverage. For Guava’s business audience, the useful question is practical: which controls deserve attention before the next security incident?
What the latest report actually says
In Detecting and countering misuse of AI: September 2026, Anthropic describes operations it identified and disrupted between December 2025 and August 2026. Its examples span cyber operations, surveillance, influence activity, scams, and other forms of misuse.
The company also describes Microsoft 365 token-theft behaviours and cases involving sensitive information passed through AI services. These are findings from Anthropic’s investigations, not a measurement of how frequently every business encounters these threats. Anthropic explicitly says its examples represent notable activity rather than typical misuse.
Guava’s assessment: the report strengthens the case for reviewing existing security controls and AI usage together. It does not establish that your organisation has been breached, or that you need to replace your technology stack.
1. Protect the session, not just the password
A sign-in token helps an application maintain access after authentication. Protecting passwords is essential, but stolen tokens create a separate risk. Microsoft recommends a layered approach to protecting tokens in Microsoft Entra ID, including endpoint hardening, access controls, detection, and measures against token replay.
Ask your IT team to document how it protects administrator accounts, identifies unmanaged devices, and investigates suspicious sign-ins. Keep MFA enabled. Review the authentication methods in use and establish a staged plan for stronger controls, with an emergency-access procedure to prevent accidental lockouts.
Microsoft’s Token Protection documentation describes device-bound sign-in session tokens designed to reduce replay risk. This is a control to assess against supported devices, applications, resources, and licensing—not a universal switch that protects every session. Test compatibility before enforcing a policy across the business.
2. Know where your business data goes
Start with an inventory of approved AI tools, browser extensions, integrations, and model-routing services. For each one, identify the business owner, the information it can access, and the terms governing data handling.
Our recommendation is to make the rules specific enough to use during a normal workday:
- Define which tools are approved for public, internal, and confidential information.
- Prohibit pasting passwords, access keys, and live configuration secrets into prompts.
- Review connectors and permissions before granting an AI tool access to shared documents or email.
- Provide an easy way for staff to request a useful tool and report accidental disclosure.
A written policy that employees cannot follow is unlikely to help. Pair the rules with approved alternatives and short examples drawn from your actual workflows.
3. Practise the first hour of an incident
Run a short tabletop exercise: an employee reports an unexpected sign-in, followed by suspicious mailbox activity. Who investigates? Who can contain access? Where are the relevant logs, and who decides whether clients must be contacted?
Have the technical team rehearse account containment, session revocation, device investigation, and recovery using its established response procedures. Confirm what each action does and does not revoke. Assign an owner to every gap the exercise reveals.
For business leaders, the deliverable is a concise response plan with named responsibilities—not another dashboard that nobody monitors.
A practical plan for the next 30 days
- Week 1: inventory AI tools, privileged accounts, and device-management coverage.
- Week 2: review sign-in policies, endpoint protection, and monitoring responsibilities with your IT provider.
- Week 3: pilot the highest-priority improvements and publish clear AI data-handling rules.
- Week 4: run the incident exercise and agree on the next actions, owners, and deadlines.
Before purchasing additional licences, compare the capabilities you already own with the controls you actually need. Microsoft notes that its token-protection recommendations span products with different licensing requirements; confirm entitlement and suitability for your environment.
Turn the headlines into a focused security review
Guava Systems helps businesses make practical Microsoft cloud decisions. If this report raises questions about your environment, start with our Microsoft 365 security-controls guide, or book a cloud assessment to identify priorities for your organisation.
The goal is measurable progress: fewer unnecessary permissions, clearer rules for AI data, and a team that knows what to do when something goes wrong.
Editorial note: Researched on 11 September 2026. This article combines attributed vendor findings with Guava’s practical recommendations. Product availability, licensing, and supported configurations can change; consult the linked Microsoft documentation before deployment.
