The Week in Tech: AI Agents, Urgent Patching and the New Reality of Automated Risk

The past seven days have made one thing clear: artificial intelligence is moving deeper into everyday business operations, while attackers are using the same automation to scale fraud, exploit vulnerabilities and create new pressure on IT teams.

For Australian businesses, the practical message is not to chase every headline or deploy every new AI feature. It is to improve visibility, tighten identity controls, patch quickly and make sure automation operates within clear limits.

AI agents are becoming an operational and financial control issue

A major theme this week has been the growing use of AI agents that can take action across business systems rather than simply answer questions. These agents can monitor applications, retrieve information, create records, send messages, call APIs and trigger multi-step workflows.

That creates obvious opportunities for productivity. A small business could use an agent to triage service requests, prepare a sales report, reconcile routine information or help staff find answers across internal documents. IT teams could use automation to identify alerts, summarise incidents and recommend standard fixes.

However, recent industry reporting has highlighted a less obvious risk: poorly controlled agents can consume far more cloud resources than expected. A single request may trigger dozens or hundreds of downstream operations, including model calls, database queries, API requests and document searches. Without limits, the result can be an unexpectedly large cloud or AI bill, as well as a much wider attack surface.

The issue is especially relevant to organisations that are experimenting with AI through software subscriptions. Teams may activate an agent inside a CRM, collaboration platform, help desk, finance system or development environment without fully understanding what data it can access or what actions it can take.

A useful industry analysis of agentic AI costs explains why governance must include cost controls, not just privacy and security controls. The business risk includes budget overruns, accidental changes to records, unauthorised data access and operational disruption.

Why it matters for businesses

For managers, this is an important shift in how technology should be approved. Traditional software reviews usually ask what an application stores and who can log in. Agent reviews must also ask what the system can do on a user’s behalf, which tools it can call and whether its activity can be stopped quickly.

An AI agent that can read a shared drive is one type of risk. An agent that can read the drive, update a customer record and send an external email is a different class of operational tool.

Practical steps

• Create an inventory of approved AI tools, agents and automated workflows across the business.
• Record what data each agent can access and which actions it is allowed to perform.
• Set spending limits, usage alerts and rate limits for AI services and cloud APIs.
• Start with read-only or approval-based workflows before allowing autonomous changes.
• Require logging for agent activity, including the user, prompt, data source, action and result.
• Review agent permissions at least quarterly and remove access that is no longer needed.

AI-powered cybercrime is making compromised email accounts more dangerous

Microsoft this week described the disruption of EvilTokens, an AI-enabled cybercrime service linked to more than 12,000 compromised email inboxes across more than 10,000 organisations worldwide. The activity affected a wide range of sectors, including construction, wholesale distribution, financial services, real estate, higher education and healthcare. Australia was among the countries with a high concentration of observed victim activity.

According to Microsoft’s Digital Crimes Unit, the service did more than collect usernames and passwords. It helped criminals analyse compromised inboxes, identify trusted relationships, find payment discussions and select people who were likely to approve transactions. It could also assist with impersonation and fraud planning.

Microsoft’s technical analysis says the service abused device code authentication flows. Victims were tricked into entering a code or completing a sign-in process on a legitimate-looking page. Attackers then used stolen tokens to access email and maintain persistence. In some cases, malicious inbox rules helped hide activity from the account owner.

This is a significant development for small and mid-sized organisations because it shows how quickly an ordinary email compromise can become a targeted financial fraud operation. The attacker does not necessarily need to understand the business manually. Automation can search the mailbox, map relationships and highlight the most valuable conversations.

The risk is not limited to Microsoft 365. The same basic pattern applies to any cloud identity platform where a stolen session, token or authorised application can provide access without the attacker needing to repeatedly enter a password.

Why it matters for businesses

Many organisations still treat MFA as the finish line for identity security. It is not. Traditional MFA can be bypassed through adversary-in-the-middle phishing, session theft, device code abuse and approval fatigue.

Business email compromise also remains a process problem. Even with strong technical controls, an attacker may succeed if staff can change bank details, approve urgent payments or release sensitive documents based on an email alone.

Practical steps

• Prefer phishing-resistant MFA, such as passkeys or hardware security keys, for administrators, finance staff and executives.
• Restrict or disable device code authentication where it is not required for normal business operations.
• Monitor for suspicious inbox rules, unusual OAuth grants, unfamiliar devices and abnormal sign-in locations.
• Revoke active sessions and tokens when an account is suspected of compromise, not just the password.
• Require an independent verification step for payment changes, urgent transfers and new supplier details.
• Train staff to treat unexpected sign-in codes, shared document invitations and urgent finance requests as potential attacks.

Patch management is now a business continuity responsibility

This week also brought renewed evidence that internet-facing vulnerabilities can move quickly from technical disclosure to operational crisis. CISA added a Microsoft SharePoint vulnerability to its Known Exploited Vulnerabilities catalogue after exploitation was observed, while security reporting also highlighted critical flaws affecting Citrix NetScaler and other widely used infrastructure products.

At the same time, Kiteworks, the secure file transfer provider formerly known as Accellion, told customers to shut down servers temporarily after receiving credible threat intelligence about a possible imminent attack. The company said the warning was precautionary and that it had no evidence its systems or customer systems had already been compromised.

The TechCrunch report on the Kiteworks warning illustrates the difficult decisions businesses may face when a critical supplier is under threat. Shutting down a file transfer service can affect customers, staff, suppliers and revenue. Keeping it online without understanding the risk can expose sensitive information.

The lesson is not that every vulnerability requires an emergency shutdown. The lesson is that organisations need a clear process for deciding what to do when a critical system is threatened.

For Australian organisations, this connects directly to operational resilience, privacy obligations, contractual commitments and customer expectations. A patching programme should not be measured only by the number of updates installed. It should also consider whether the organisation can identify exposed assets, isolate a vulnerable service, maintain a backup process and communicate with affected stakeholders.

Why it matters for businesses

Attackers often target products that sit at the edge of the network or handle valuable information. These include remote access systems, file transfer platforms, content management systems, VPNs, firewalls, collaboration tools and business applications.

A vulnerability in one of these systems can affect the whole business even if internal endpoints are well protected. A single unpatched server may provide an entry point for data theft, ransomware, credential harvesting or disruption.

Practical steps

• Maintain an accurate inventory of internet-facing systems, including systems managed by suppliers or third parties.
• Prioritise vulnerabilities listed in active exploitation catalogues and vendor emergency advisories.
• Define patching deadlines based on exposure, exploitability, business criticality and available mitigations.
• Test emergency isolation procedures for remote access, file transfer, VPN and identity systems.
• Keep offline or logically isolated backups and regularly test restoration rather than assuming backups will work.
• Establish an escalation process that gives IT authority to take a service offline when the risk is severe.

Mobile and endpoint updates cannot be left to individual users

Apple security updates were another reminder this week that employee devices are part of the corporate attack surface. Security researchers and industry reporting pointed to emergency patches for a CoreGraphics vulnerability affecting Apple operating systems, with evidence that the issue may have been exploited in targeted attacks.

The affected component handles graphics and image processing. Vulnerabilities in these areas can be triggered through specially crafted content, which means a user may not need to download an obvious executable for a device to be exposed.

The Apple security releases page remains the authoritative source for current platform updates and security content. The practical concern for businesses is not limited to whether an employee owns an iPhone, iPad or Mac. It includes whether those devices access corporate email, customer systems, cloud storage, finance platforms or internal applications.

Bring-your-own-device policies can make this difficult. If an organisation cannot see the operating system version, enforce screen locks, require encryption or remove access from a lost device, it may not have enough control to manage the risk.

Why it matters for businesses

Endpoint security is often strongest on company-issued Windows laptops and weakest on mobile devices and personally owned equipment. That gap is attractive to attackers because phones are used for MFA, email, messaging and password recovery.

A delay of several days may be understandable for a personal device. It is harder to justify when the device is used by an administrator or finance employee to access sensitive systems.

Practical steps

• Use mobile device management or mobile application management for devices that access business data.
• Set a minimum supported operating system version and block access from devices that fall below it.
• Enable automatic updates wherever practical and verify compliance through central reporting.
• Require encryption, screen locks, remote wipe capability and secure backups.
• Apply stronger controls to privileged users and staff who approve payments or handle personal information.
• Review whether business applications really need to be available on unmanaged personal devices.

The security workforce still needs better automation, not blind automation

Another recurring story this week has been the pressure on security teams. Threat volumes are increasing, while many organisations still rely on fragmented tools, manual investigation and limited specialist capacity.

AI can help. Modern security platforms can summarise incidents, correlate alerts, analyse suspicious files and suggest response actions. Microsoft and other major vendors are increasingly positioning security agents as a way to reduce repetitive work and help smaller teams operate more effectively.

But the most credible guidance this week has focused on supervised automation. Industry experts continue to warn that AI should not be treated as a replacement for judgement, context and accountability. The best model for many organisations is human-on-the-loop security, where automation handles routine analysis but a qualified person remains responsible for high-impact decisions.

The recent discussion of human-led oversight in managed security is particularly relevant to smaller businesses. An organisation may not have the budget for a large internal security operations centre, but it can still combine managed detection, clear escalation paths and well-defined response playbooks.

Why it matters for businesses

Automation can reduce the time required to investigate an alert, but it can also amplify a mistake. If an automated system disables an important account, blocks a supplier, deletes a file or isolates a server incorrectly, the business may suffer operational damage.

The answer is to automate repeatable, reversible tasks first. Use automation to collect evidence, enrich alerts, open tickets and recommend next steps. Require approval before actions that affect customers, payments, production systems or critical infrastructure.

Practical steps

• Identify the most repetitive security tasks and automate those before attempting fully autonomous response.
• Create approval gates for account disabling, data deletion, payment-related actions and production changes.
• Test AI-generated recommendations against known scenarios before relying on them in live incidents.
• Use an external managed security provider if internal staffing cannot provide continuous monitoring.
• Keep incident playbooks current and make sure staff know who has authority to make urgent decisions.
• Measure automation by outcomes such as response time, false positives and avoided downtime, not by novelty.

What This Means For Your Business

The news from the past week points to a consistent business reality. AI is increasing the speed of useful work, but it is also increasing the speed of fraud, exploitation and operational mistakes. The organisations that benefit most will not necessarily be those with the biggest technology budgets. They will be those that establish sensible controls before automation becomes deeply embedded.

Start with visibility. Know which AI tools, cloud services, mobile devices, suppliers and internet-facing systems your organisation depends on. Then identify the systems that could create the greatest financial, legal or reputational impact if compromised or misconfigured.

Next, strengthen the basics that continue to matter: phishing-resistant MFA, reliable patching, tested backups, monitored identity activity, independent payment verification and clear incident response responsibilities. These measures are not outdated because AI is developing. They are becoming more important because AI makes successful attacks faster and more targeted.

Finally, treat automation as a controlled business capability. Give agents narrow permissions, set spending limits, log their activity and keep people involved when decisions are difficult to reverse. A small number of practical improvements each week can do more for resilience than a large technology project with no clear owner.

Call Webwire on 08 9386 0053 or contact us at enquiries@webwire.com.au.