⚡ Quick Verdict: Securing AI Under the Essential Eight
Securing Generative AI under the ACSC Essential Eight requires strict enforcement of Control 1 (Application Control) and Control 5 (Restrict Administrative Privileges).
Because employees are increasingly bypassing perimeter security using unauthorized AI tools (“Shadow AI”), IT directors must use Cloud Access Security Brokers (CASBs) and DNS sinkholing to block unapproved LLM endpoints. Furthermore, custom AI agents (like Copilot or Chatbase webhooks) must now be classified as Non-Human Identities (NHIs) and restricted using Conditional Access policies in Microsoft Entra ID or Okta.
Note: While the ASD announced the transition to the new “Essentials Series” in June 2026, Maturity Level 2 (ML2) remains the mandatory baseline for Australian organizations over the next 24 months.
Do not abandon your current E8 audits. The ASD has explicitly stated that the transition period to the new Essentials for Enterprise IT chapter will take up to 24 months. During this window, achieving Maturity Level 2 (ML2) across the original controls—while mapping them to the new June 2026 ISM GenAI requirements—remains the mandatory benchmark for securing government contracts, DISP memberships, and passing cyber insurance audits.
The June 2026 Update: From “Essential Eight” to the “Essentials Series”
For years, the Essential Eight (E8) has served as the de facto cybersecurity baseline for Australian businesses and government contractors. However, the rapid proliferation of Generative AI, microservices, and Bring-Your-Own-Device (BYOD) policies rendered the 8-year-old framework obsolete against modern threats.
On June 24, 2026, the Australian Signals Directorate (ASD) announced a massive overhaul, initiating a transition from the prescriptive Essential Eight to a broader, threat-informed framework called the “Essentials Series”.
What this means for IT Directors:
Do not abandon your current E8 audits. The ASD has explicitly stated that the transition period will take up to 24 months. During this window, achieving Maturity Level 2 (ML2) across the original eight controls remains the mandatory benchmark for securing government contracts and passing cyber insurance audits.
However, your compliance strategy must now account for AI. If your current Application Control policy blocks unauthorized .exe files but allows employees to paste proprietary code into a public web browser running Claude or ChatGPT, you will fail your next audit.
Control 1 (Application Control): Destroying “Shadow AI”
“Shadow AI” occurs when employees secretly feed sensitive company data, source code, or legal contracts into unapproved, consumer-grade AI tools. Because these public LLMs use user inputs as training data, Shadow AI represents a catastrophic data exfiltration risk.
To achieve ML2 Application Control in an AI-driven environment, traditional endpoint antivirus is not enough. You must block the AI at the identity and DNS layers.
1.Deploy a Cloud Access Security Broker (CASB):
Integrate a CASB (like Microsoft Defender for Cloud Apps) to monitor your network traffic. CASBs can automatically identify and tag over 400+ consumer Generative AI applications currently operating in the wild.
2.Implement DNS Sinkholing:Blocks browser-based AI access.
Once the unapproved AI tools are identified, configure your network edge (via Cisco Umbrella or Zscaler) to sinkhole the DNS requests. If an employee tries to visit an unapproved LLM, the network drops the connection before the page loads.
3.Whitelist Approved Vendor AI:
Explicitly approve secure, enterprise-grade AI applications. For example, if you use a closed-loop system like Chatbase or Tidio for customer support, whitelist their specific API endpoints to ensure your business operations continue uninterrupted.
Interact with the flowchart below to understand the logical routing required to block unauthorized AI identities at the network edge:
The Cultural Fix: Give Them a Secure Alternative
Shadow AI is rarely an act of employee rebellion; it is a sign that your staff is eager to innovate and be productive. If you block consumer LLMs via DNS sinkholing but fail to provide a secure alternative, employees will simply find workarounds (like using AI on their personal phones off-network).
True Application Control means balancing restriction with enablement. You must replace the blocked Shadow AI with a managed solution—like a secure Microsoft 365 Copilot license or a privately hosted Chatbase/CustomGPT workspace—where company data is protected by enterprise DPAs and zero-retention policies.
Shadow AI Blocking Infrastructure
⚡ Shadow AI Identity & Network Blocking Logic
Adjust the parameters to see how Cloud Identity and CASB infrastructure routes or blocks AI requests.
The June 2026 ISM AI Controls: Securing “OFFICIAL: Sensitive” Data
While the Essential Eight is transitioning into the Essentials for Enterprise IT, the underlying Information Security Manual (ISM) was just updated in June 2026 to introduce the first-ever explicit technical controls for Artificial Intelligence. If you need to build this, see our technical guide: How to Build a Chatbase Human Handoff Webhook for High-Risk Queries.
If your AI chatbot or internal agent handles data classified as “OFFICIAL: Sensitive,” you will fail your IRAP (Information Security Registered Assessors Program) assessment unless you comply with these new mandates:
- ISM-2112 (Disable External Fetching): If your custom AI processes sensitive internal data, its ability to autonomously query the open internet must be completely disabled. This prevents the AI from accidentally leaking private data into public search indexes.
- ISM-2113 (Human-in-the-Loop): AI applications must be configured to flag organizationally defined “risky actions” for human approval before they execute. (For example, an AI cannot automatically issue a $5,000 refund; it must ping a human manager first).
- ISM-2114 (Baseline Deviation): You must actively monitor your AI’s behavior. If the bot suddenly starts hallucinating or giving unauthorized advice, your monitoring tools must instantly kill its access token.
The Fix: You cannot achieve ISM-2113 compliance with a basic ChatGPT wrapper. You must use a platform capable of pausing workflows and routing to live agents.
(If you need to build this, see our technical guide: How to Build a Chatbase Human Handoff Webhook for High-Risk Queries.)
Control 5 (Restrict Admin Privileges): The “Non-Human Identity” Threat
The highest-risk vulnerability in 2026 is no longer a phishing email—it is a compromised API key belonging to an AI agent.
When you connect an AI agent (like Microsoft 365 Copilot or a custom Zapier webhook) to your CRM, that AI is granted access to read emails, modify files, and trigger workflows. In cybersecurity terms, this AI is a Non-Human Identity (NHI). If an attacker injects a malicious prompt into your AI agent, they can hijack its permissions to exfiltrate your entire database.
To comply with Control 5 for AI, you must:
- Treat AI as High-Risk Administrators: Never grant an AI agent “global read/write” access. Apply the Principle of Least Privilege (PoLP). If your AI is built to check order statuses, restrict its API access strictly to the shipping database.
- Enforce Conditional Access for Agents: Utilizing mid-2026 features within Microsoft Entra ID and Okta Workflows, you can now apply Conditional Access policies directly to machine identities. Configure your identity provider to instantly revoke the AI’s access token if it attempts to pull data from outside its designated geo-location or outside standard business hours.
Control 4 (User Application Hardening): The Browser Extension Danger
Many employees try to bypass DNS blocks by installing AI-powered Chrome extensions (like grammar checkers, PDF summarizers, or email drafters). These extensions require permissions to “Read and change all your data on all websites.”
If an employee opens a highly classified legal document in their browser while a rogue AI extension is active, the extension instantly scrapes the text and sends it to a foreign server.
The Mitigation Strategy:
Under Control 4 (User Application Hardening), you must implement managed browser policies via Microsoft Intune or Google Workspace.
- Block the installation of all unverified third-party web extensions.
- Disable the execution of Java and web advertisements on high-risk machines.
- Enforce a strict “Allowlist” only for pre-vetted, SOC2-compliant productivity extensions.
Control 8 (Daily Backups): Protecting Your AI’s “Memory”
The final Essential Eight strategy requires Daily Backups of critical data. In an AI context, your chatbot’s intelligence relies entirely on its Vector Database (the parsed data chunks that feed its RAG architecture).
If a malicious actor successfully executes a “Prompt Injection” attack and poisons your AI’s underlying knowledge base, the bot may start generating hallucinations, misinformation, or offensive content to your customers.
Under ML2, you must ensure that your AI vendor (or internal server) conducts isolated, immutable daily backups of your clean Vector Data. This ensures that if the active AI database is corrupted, you can instantly roll back to a clean snapshot within your required Recovery Time Objective (RTO).
Conclusion: Bridging the Gap to 2028
The ASD’s transition to the Essentials Series over the next 24 months acknowledges a hard truth: perimeter security is dead. In an era of AI agents and cloud microservices, your identity provider (Entra ID, Okta) is your new firewall.
By applying strict Application Control to block Shadow AI, treating your custom AI agents as high-risk Non-Human Identities, and locking down web browsers, you will not only satisfy your current Essential Eight ML2 requirements but successfully future-proof your organization for the upcoming Essentials Series framework.
Ensure your infrastructure complies with the latest OAIC regulations by reading our guide: Are Your AI Chatbots Breaking Australian Privacy Laws?
Frequently Asked Questions
Does Microsoft 365 Copilot automatically meet the Essential Eight requirements?
No. While Microsoft 365 Copilot operates within your secure tenant, simply buying a license does not make you compliant. To meet Maturity Level 2, IT administrators must manually configure sensitivity labels (Purview) and restrict the Copilot bot’s administrative privileges (Control 5) so it cannot access unauthorized SharePoint files.
What is the deadline to transition from the Essential Eight to the new Essentials Series?
The Australian Signals Directorate (ASD) announced the transition to the new Essentials for Enterprise IT framework in June 2026. However, organizations have a 24-month grace period. Until June 2028, Maturity Level 2 (ML2) of the original Essential Eight remains the mandatory compliance baseline for government contracts.
Can we use a public AI like ChatGPT if we have a BYOD (Bring Your Own Device) policy?
Under Control 1 (Application Control), allowing employees to use public LLMs on unmanaged personal devices to process company data is a severe compliance violation. You must route all BYOD traffic through a CASB or VPN that blocks consumer AI endpoints, forcing staff to use approved, enterprise-secured AI tools.
