The Week in Tech: AI Agents, Hidden Access Risks and a New Reality for IT Leaders

Artificial intelligence is moving quickly from a productivity experiment to an active participant in business operations. This week’s technology and cybersecurity news shows why organisations need to treat AI adoption, remote access and software supply chains as connected business risks.

For small and mid-sized organisations, the message is practical rather than theoretical. You do not need a large security operations centre to respond well, but you do need clear ownership, disciplined access controls and a willingness to review how technology is actually being used.

1. Microsoft warns that AI agents can become insider threats

Microsoft used a security briefing published on 19 October 2026 to highlight a growing concern: AI agents deployed inside organisations should not automatically be treated as trusted assistants. Agents may be given access to company data, business applications, email, code repositories and financial workflows. If they misunderstand instructions, follow malicious content or operate with excessive permissions, the resulting incident may look very similar to an insider event.

The warning comes as vendors develop controls designed to adjust an agent’s permissions as its behaviour or risk level changes. The broader trend is clear. Security teams are beginning to monitor AI agents as identities with access privileges, rather than viewing them as simple software features.

According to Microsoft’s security guidance, frontier AI models should not be assumed to understand an organisation’s rules or exercise human judgement consistently. That does not mean businesses should stop using AI. It means AI needs the same governance expected of employees, contractors and service accounts.

Why it matters for businesses

A small business might connect an AI assistant to a shared drive, customer relationship platform or accounting system without realising how much authority it has. A poorly scoped agent could expose confidential information, send an unauthorised message, change a record or approve an action that would normally require human review.

The financial and reputational impact can be significant. A data breach involving an AI tool may still trigger privacy obligations, customer notifications, contractual reporting requirements and insurance questions. Australian businesses should also consider how AI use interacts with privacy expectations, industry regulations and existing risk management policies.

Practical steps to take

• Create an inventory of every AI assistant, agent and automation connected to business systems.
• Give each agent a named business owner, a documented purpose and a defined review date.
• Use separate identities and least-privilege permissions instead of shared administrator accounts.
• Require human approval for payments, changes to customer records, external communications and access to sensitive data.
• Log agent activity and review unusual actions, repeated failures and attempts to access unrelated systems.

2. LiteLLM vulnerabilities show why AI infrastructure belongs in the patch cycle

Security researchers disclosed serious vulnerabilities in LiteLLM, an open source gateway used to connect applications with multiple AI providers. Reports published on 14 and 15 October described flaws that could allow unauthorised users to gain administrative control, access sensitive files, execute code or steal AI provider credentials, depending on the affected configuration and version.

One of the reported issues involved server-side request forgery, which can cause a service to send requests to destinations controlled by an attacker or to internal services that should not be publicly reachable. Other weaknesses affected authentication and administrative functions. The practical concern is that an AI gateway may hold valuable API keys while also seeing prompts, responses and business data moving through the system.

The issue was covered by specialist security reporting and technical industry coverage. Businesses should verify the official project advisories and package release notes rather than relying on a general assumption that a cloud-hosted AI service is automatically secure.

Why it matters for businesses

Many smaller organisations are adopting open source AI tools through developers, consultants or departments working independently. That can create an untracked layer of infrastructure outside the normal endpoint, server and cloud management processes.

If an AI gateway is compromised, the attacker may be able to use stolen keys to run up provider charges, access confidential prompts, impersonate internal applications or pivot into other systems. The consequences may include unexpected bills, intellectual property loss and a difficult investigation into what information was sent through the platform.

Practical steps to take

• Ask your IT team or managed service provider whether LiteLLM or another AI gateway is in use anywhere in the environment.
• Identify the exact version, exposure level and authentication configuration for every deployment.
• Apply the vendor’s fixed release or mitigation guidance as a priority, especially for internet-facing systems.
• Rotate AI provider keys if there is any possibility that the gateway was exposed or accessed without authorisation.
• Store credentials in a managed secrets vault, restrict outbound network access and monitor unusual API usage.

This is a useful reminder that AI security is not limited to model behaviour. The surrounding gateways, plugins, connectors, scripts and cloud permissions deserve the same scrutiny as conventional business applications.

3. North Korean remote worker schemes put hiring and access controls in the spotlight

US authorities and law enforcement partners disrupted a North Korean remote IT worker operation during a series of actions reported in October. The investigation involved more than 40 suspected laptop farms in Texas, the seizure of nearly 187 laptops and charges involving individuals alleged to have helped overseas workers appear to be based in the United States.

The scheme relied on stolen or fabricated identities, remote access and domestic facilitators who hosted company-issued equipment. The operators were able to obtain employment with legitimate organisations, creating the possibility of access to source code, customer data, internal systems and intellectual property.

The US Department of Justice has previously warned that these schemes can involve the use of US-based laptops and intermediaries to conceal the true location of workers. Recent reporting from BleepingComputer shows how the issue is moving from a fraud concern into a broader insider risk and supply chain problem.

Why it matters for businesses

The lesson is not that remote work is unsafe. The lesson is that identity verification cannot stop at a video interview and a signed contract. A business may employ a genuine-looking candidate while someone else performs the work, controls the device or accesses systems from an undisclosed country.

This matters especially to organisations that outsource software development, helpdesk services, data entry or cloud administration. It may also affect suppliers and subcontractors further down the chain. In Australia, businesses should consider whether their vendor management processes, privacy agreements and data residency requirements adequately address subcontracted access.

Practical steps to take

• Use a structured identity verification process before issuing equipment, credentials or privileged access.
• Confirm that the person interviewed is the same person using the account and device during onboarding.
• Ship equipment only through approved channels and record device ownership, location and custody.
• Restrict access for new starters until role, identity and work location checks are complete.
• Review remote access logs for impossible travel, unusual time zones, unexpected VPN use and simultaneous sessions.

For higher-risk roles, consider a staged access model. Begin with a limited environment, require supervised work for sensitive tasks and expand access only after behaviour matches the agreed employment and location details.

4. The market is building tools to govern employee use of AI

On 14 October, Harmonic Security announced a 50 million US dollar Series B investment for technology designed to classify and control workplace AI traffic. The company says its platform can assess how employees use browser-based AI tools, desktop applications and agentic workflows, then apply policies in real time.

The funding is one sign of a larger market shift. Businesses are no longer debating whether staff will use generative AI. They are looking for ways to understand which tools are being used, what information is being shared and where the organisation needs to permit, restrict or monitor activity.

Coverage from SiliconANGLE and VentureBeat described the company’s approach to workplace AI governance. Businesses do not necessarily need to buy a specialist platform immediately, but they do need a usable policy and a way to enforce it.

Why it matters for businesses

A blanket ban on AI is unlikely to work. Employees may still use personal accounts or unapproved tools to summarise documents, write code, analyse data or prepare customer communications. That creates shadow AI, where the organisation has little visibility over data handling, retention or vendor terms.

A clear and practical policy can reduce this risk while preserving productivity. Staff should know which tools are approved, what information must not be uploaded and when human review is required. Managers also need a process for requesting new tools so useful innovation does not become an uncontrolled exception.

Practical steps to take

• Publish a short AI use policy written in plain language and linked to existing privacy and acceptable-use rules.
• Classify information into simple categories such as public, internal, confidential and highly restricted.
• Prohibit the use of personal AI accounts for confidential customer, employee or commercial information.
• Approve a small set of business tools with suitable security, retention and administration controls.
• Review AI usage through identity logs, browser controls, data loss prevention or periodic staff surveys.

Training is essential. Employees need examples of safe and unsafe prompts, including the risks of uploading contracts, source code, customer records, passwords or personal information.

5. AI is increasing the scale and speed of attacks

A report published this week described an incident in which an AI-assisted operator attempted to break into more than 1,200 websites during the week ending 18 October. Reporting indicated that the activity used agentic tools to automate reconnaissance and testing across a large number of targets.

The significance is not that every AI-assisted attempt will succeed. It is that attackers can now test more organisations in less time, with fewer people and lower operating costs. A campaign that once required a team of specialists may be partially automated, allowing criminals to search for weak passwords, exposed services, vulnerable applications and poorly configured cloud resources at scale.

The incident was covered by Tom’s Hardware. Businesses should treat the report as a reminder to improve basic resilience rather than as a reason to chase every new AI security product.

Why it matters for businesses

Small organisations are attractive targets because they often have internet-facing systems but limited monitoring and slow patching. Automated reconnaissance makes it easier for attackers to identify those weaknesses without specifically researching the company in advance.

The same dynamic applies to phishing and business email compromise. AI can help criminals produce convincing messages, adapt language to local audiences and respond to victims in real time. A polished email is no longer evidence that a request is genuine.

Practical steps to take

• Turn on phishing-resistant MFA where possible, especially for administrators, finance staff and remote access.
• Maintain an accurate list of internet-facing systems and remove services that are no longer required.
• Prioritise critical patches for public-facing applications, VPNs, firewalls, email systems and identity platforms.
• Configure alerts for impossible travel, unusual login locations, mass downloads and new forwarding rules.
• Test payment-change and supplier-bank-detail procedures using an independent verification step.

Businesses should also rehearse what happens after a suspicious login or ransomware alert. Knowing who can disable an account, isolate a device, contact the bank and notify customers can reduce decision time when pressure is highest.

What This Means For Your Business

The common thread across this week’s news is that trust must become more specific. Trust should not be granted simply because a request comes from an employee, an AI assistant, a contractor, a software package or a familiar cloud service. Access needs to match the task, the person, the device and the current level of risk.

For most small and mid-sized organisations, the best response is a focused 30-day review. Start by listing your AI tools, remote workers, critical suppliers, privileged accounts and internet-facing services. Then confirm that each has an owner, an appropriate level of access and a recovery plan.

Next, close the basics that attackers continue to exploit. Patch exposed systems, enforce MFA, remove dormant accounts, rotate old credentials, improve logging and train staff to verify unusual requests. These actions are affordable compared with the cost of a serious breach and they support both operational resilience and customer confidence.

AI can still deliver real value through faster service, better analysis and more efficient administration. The organisations most likely to benefit will be those that treat governance as an enabler, not a brake. Build sensible controls now, review them regularly and make sure every new automation has a clear business purpose and a human owner.

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