Cybersecurity Awareness

Min read

Your Next Insider Threat Might Be an AI Agent

Published on

Oct 19, 2026

Your Next Insider Threat Might Be an AI Agent

AI agents can read email, move files and run code with the permissions they are given. That makes them useful, and a new kind of identity your business must manage.

A developer asked an AI coding agent to clean up an old project copy. According to the developer’s account, the agent followed directory links into a live project and deleted 48,218 files in 103 seconds.The story was reported by TechRadar in September 2026; it has not been independently verified. Even with that caveat, the failure is familiar: the agent could reach more than the task required, and no checkpoint stopped a destructive action.

The lesson is not that AI agents are malicious. It is that software acting on a person’s behalf can make mistakes, follow hostile instructions or be misused, and can do so at machine speed. If an agent can access business data or change systems, it needs the same deliberate oversight as any other identity.

Why AI agents change the security picture

A chatbot responds with text. An AI agent can also use tools: search a mailbox, update a customer record, share a document or execute code. Its reach depends on the account, tokens and permissions connected to it. If an employee connects an agent to a personal work account, the agent may inherit access intended for that employee, along with the ability to take actions the employee could take.

Agents also process information they do not control. An email, web page or document can contain instructions designed to manipulate anAI system into ignoring its intended task. This is commonly called prompt injection. If the agent can both read sensitive information and send it elsewhere, a malicious instruction may turn a routine workflow into a path for data exposure.

The scale of adoption is moving faster than some controls. In Palo Alto Networks’ 2026 Identity Security Landscape report, only 37% of surveyed organizations said they could revoke an AI agent’s credentials,and 30% reported immutable audit logging for agent activity. Gartner has separately forecast that 25% of enterprise breaches will be traced to AI agent abuse by 2028. These are signals to prepare, not proof that every organization faces the same level of risk.

Four AI agent security risks to manage

1. Excessive access

Teams often grant broad permissions to get a new tool working quickly. If the agent only needs to summarize one project folder, it should not also be able to search the finance drive, delete records or share files externally. Start with the narrowest access that lets it do its job.

2. Prompt injection

Instructions hidden in content the agent reads can try toredirect its behavior. Treat emails, documents and web pages as untrustedinput. Limit which tools the agent can use, separate reading from high-impactactions, and require approval before it sends data or changes importantrecords.

3. Unclear activity trails

When an agent operates through a person’s account, its actions can blend into ordinary user activity. Give each business agent a distinct identity where possible, record what it accesses and changes, and make sure someone can review those logs. Clear attribution makes investigation and access revocation faster.

4. Shadow agents

Employees may connect browser extensions, meeting assistants or automation tools to company accounts without involving IT. The result is an unknown set of integrations and permissions. Make it easy to request approved tools, and review connected applications regularly so staff are not pushed toward workarounds.

A practical 30-day plan for small businesses

Week 1: Find the agents

List AI tools, extensions and automations connected to email, file storage, CRM, collaboration and code repositories. Record who owns each one, which account it uses, what data it can reach and what actions it can take. Check both approved products and user-authorized applications.

Week 2: Reduce permissions

Remove integrations with no clear owner or business purpose. For the rest, replace broad access with specific folders, read-only access or limited scopes wherever possible. Use a dedicated identity for each agent rather than a shared or personal login, and document how to revoke its credentials.

Week 3: Add approval and monitoring

Require a person to approve high-impact actions such as deleting data, changing permissions, making payments or sharing files outside the organization. Enable activity logging and alerts for unusual behavior,including bulk downloads, repeated failed actions and access outside the agent’s normal role.

Week 4: Set the rules

Publish an AI acceptable use policy that explains which tools are approved, what data may be used, who can authorize new agents and which actions require human review. You don’t have to start from a blank page: our free AI Acceptable Use Policy template covers approved tools, what should never be pasted into AI and a staff acknowledgement form. Brief staff on prompt injection and shadow AI. Assign an owner to review the agent inventory at least quarterly and whenever a role or vendor changes.

A five-minute check you can do today

Open the connected-app or enterprise-applications area for your company accounts. In Microsoft 365, review enterprise applications in the Entra admin center. In Google Workspace, review API controls in the Admin console. For each AI-related connection, ask: who authorized it, what can it access, and can we revoke it quickly? Remove or investigate anything nobody can explain.

Keep the productivity. Put boundaries around it.

AI agents can save time, but speed makes access mistakes and hostile instructions more consequential. Treat each agent as a distinct identity: give it an owner, limit what it can reach, log what it does and keep a person in control of consequential actions. Start by finding the agents already connected to your systems; you cannot manage access you do not know exists.

Want a head start? Download our free cybersecurity guides, including the AI Acceptable Use Policy template, an SME Cyber Readiness Checklist and an MFA setup guide. No sign-up needed. If you would like help reviewing the AI tools connected to your systems, talk to our team.

‍

Other Articles

Cyberthreats

Persistent Siege: Large-Scale Brute Force Attacks on Cisco ASA SSL WebVPN

Since December 2023, we've been tracking significant brute force campaigns targeting Cisco ASA SSL WebVPN. Daily attacks range from 1,000 to 8,000 attempts, originating from residential proxies, VPNs, and suspicious hosting providers. Attackers use automated scripts and common credentials to breach systems. As of June 4, 2024, these attacks continue.

GDPR, Google Analytics and Matamo

GDPR is the world’s strongest set of data protection rules, improving how people access information about themselves and limiting what organizations can do with personal data.

New Mirai Variants Targeting Latest Vulnerabilities in IoT Devices

This campaign was initially observed starting on 10/04/2024, soon after the exploit for CVE-2024-3273 became available. Subsequently, on 15/04/2024, the exploitation of CVE-2024-3721 began from the same attacker. We are confident that it is the same actor behind both exploits, as the IP addresses and malware samples were identical for both vulnerabilities.

Andromeda Botnet: A Threat to Albania’s Cybersecurity

The Andromeda botnet, also known as Gamarue, Wauchos, and Andromeda Stealer, is a sophisticated and long-standing malware family that has been used for various purposes, such as spamming, DDoS attacks, and credential theft. It has been active since 2011 and has infected millions of devices worldwide.