Three Malaysian colleagues reviewing a laptop and discussing a request

AI cyber attack risks: trust, approvals and access at work

Short answer: an AI cyber attack can use AI tools to help find weaknesses, test access and carry out parts of an intrusion. Recent reports on South Korean financial organisations and Southeast Asian targets show this happening outside a laboratory. For Malaysian workplaces, the response includes tighter control of system access, independent confirmation of sensitive requests and room for staff to pause when an approval feels wrong.

A request arrives from a familiar name. It uses the language of the business, refers to a current job and asks for something that sounds routine.

Perhaps it is a supplier changing bank details. Perhaps an assistant needs access to a shared folder to finish a report.

Those are different situations, with different risks. Both put a decision in front of someone who may be busy, under pressure and inclined to help.

The recent regional cyber incidents affecting Malaysian readers raised questions about the role of AI. Two security reports give workplaces a reason to look more closely at what their people and systems are allowed to approve.

What the South Korean attacks show

CrowdStrike’s 7 October report describes a campaign against South Korean financial organisations from late September to early October 2026 that resulted in data being taken.

Researchers found AI coding-session histories, configuration files and memory files on infrastructure controlled by the attacker. The material showed use of ARTEX, an agentic penetration-testing tool, alongside large language models.

That is more specific than a claim that a hacker probably used a chatbot. The recovered files gave researchers a view of the tools and activity involved.

There are limits to the finding. CrowdStrike said the number of affected organisations remained unconfirmed, and it did not attribute the activity to a named adversary.

It would be wrong to stretch that report into a claim that every South Korean bank incident had the same cause. It would be equally wrong to dismiss the use of AI in this campaign as a distant possibility.

AI cyber attack activity in Southeast Asia

In a 6 November report, Microsoft described Storm-3358, a Cambodia-based intrusion cluster that primarily targets financial organisations in Southeast Asia.

Microsoft observed the group using ARTEX with conventional automated tools to scan internet-connected systems, exploit security weaknesses and install remote-access software. Its report also described stolen credentials and compromised cloud identities being used during attacks.

The combination matters. An attacker can use AI alongside familiar intrusion methods, so a workplace cannot judge its exposure by asking only whether it has bought an AI product.

Nor does this report establish that a particular Malaysian incident was AI-driven. Incident-specific claims still need incident-specific support.

For managers, the useful question is how much damage an unwanted action could cause before someone notices. That question applies to an outsider with stolen access and to an authorised AI assistant that has been given more permission than its job needs.

Trust depends on what is being requested

People often use familiar language as a shortcut for trust. A message sounds like a colleague, so it receives less scrutiny than one full of odd phrasing.

That shortcut deserves less weight when a decision involves money, confidential information or a change in access. Good wording does not tell the recipient whether the sender has the right to request the action.

Imagine a finance executive receiving an urgent message about an overdue supplier payment. The invoice looks familiar, but the bank details have changed and the sender wants payment before the afternoon cut-off.

The executive asks an AI assistant, “Summarise this supplier email and the action requested.” In this hypothetical example, the summary accurately says that payment is urgent and gives the new details.

Nothing in that summary establishes whether the change is genuine. A useful summary can still carry an unauthorised request from the inbox to the approval desk.

  • Familiar wording: a convincing tone is a reason to read the request, not a reason to waive an approval rule.
  • A new destination: changed payment details or a new account deserve confirmation through a route the organisation already trusts.
  • A claim of urgency: a deadline may be real, but it does not settle who has authority to change the arrangement.
  • A fluent AI summary: it can describe what a message says without proving that the message should be acted on.
A woman making a phone call beside a coworker in an office
AI-generated editorial image of an employee checking a request by phone.

Approvals need an independent route

A company may already require someone to approve a payment. That control becomes weaker if the approver relies on the same unverified message as the person preparing it.

Two people reading the same forged request can reach the same wrong conclusion. The value of a second approval depends on what the second person actually checks.

For sensitive changes, an independent confirmation route gives the team something else to rely on. That might be an established supplier contact or the company’s existing change-authorisation process.

A phone number supplied only in the disputed message does not provide much independence. Neither does asking the same account to repeat its request.

These are approval principles, rather than a promise that one check will prevent every loss. Finance and security teams need controls suited to the amounts, systems and obligations involved.

The workplace culture matters too. If a manager praises speed while treating every pause as obstruction, staff have a reason to skip the questions the company says it wants them to ask.

Giving an AI assistant access creates another responsibility

An assistant that drafts a paragraph from supplied text has a different reach from one that can search company files, send messages or change records.

The difference is what the assistant can do after it receives an instruction. As access grows, a mistake or an unwanted instruction can have consequences beyond a poor answer.

Consider a hypothetical operations team connecting an assistant to a shared document collection. The intended task is to help staff find current procedures.

If the connection also allows the assistant to alter files or reach unrelated personnel records, its permissions exceed that purpose. The team has created exposure that better wording alone cannot fix.

  • Permissions: access should fit the work the assistant is expected to perform, including the difference between reading and changing information.
  • Ownership: someone should be accountable for the connection and for reviewing whether it is still needed.
  • Consequential actions: sending outside the company, changing access or moving money needs an approval boundary appropriate to the risk.
  • Withdrawal: the company needs a workable way to remove access when the job, person or service changes.

This extends the concern behind Shadow AI at work. An unapproved connection can be harder to manage than an unapproved writing tool because the company may not know what it can reach.

AI cyber attack risk graphic showing authority, payment changes, AI permissions and permission to pause
Workplace decisions around request authority, changed accounts and system permissions.

Staff need a place to take an uncertain request

Someone may notice a concern without being able to name the attack. A supplier’s explanation feels unusual, a login request arrives out of context, or an assistant asks for access unrelated to the task.

Requiring that person to prove an attack before raising the concern sets the bar too high. The team needs a known person or function that can take responsibility for the next decision.

Managers can make that easier by being clear about which actions staff may pause and who can authorise a sensitive exception. A pause should not depend on the employee’s confidence in arguing with a senior colleague.

An employee AI policy belongs in that conversation, alongside the company’s existing finance and security rules. Staff should not have to guess which rule wins when an AI-assisted task crosses into a controlled process.

Where training helps, and where technical work remains

Workplace training can help people judge an AI summary, notice a request that exceeds its purpose and recognise when approval belongs elsewhere. Practice with realistic decisions can make those boundaries less abstract.

It cannot establish that a server is secure or remove a stolen credential from an attacker’s hands. The security team still owns the technical assessment and response.

Microsoft’s report describes exploitation of system weaknesses as well as AI-assisted activity. Organisations need to maintain their technical controls while helping employees understand their own role.

The aim is proportionate responsibility. Staff can ask better questions without being expected to investigate every suspicious request themselves.

Frequently asked questions

What is an AI cyber attack?

It is a cyberattack in which AI helps with part of the attacker’s work, such as finding weaknesses or carrying out intrusion activity. The degree of automation varies, and the label does not mean a system acted without human direction.

Do the regional reports prove Malaysian banks were attacked with AI?

No. The CrowdStrike and Microsoft reports describe particular activity and targets; they should not be used to attribute an unrelated Malaysian incident.

Should companies stop using AI at work?

A company should assess the purpose, data and permissions of each use. A drafting task and an assistant with access to company systems create different exposure, so one blanket answer would hide important decisions.

Can an AI-generated summary approve a request?

A summary can help a person understand the request, but it does not establish the sender’s identity or authority. Approval still needs to follow the company’s rules for the action involved.

What should ordinary employees be responsible for?

Employees should know their approval limits, where to raise a concern and what information or access they may share. They should have support from the people responsible for finance, data protection and security when a request falls outside those limits.

A workplace that can pause

AI-assisted attacks add pressure to decisions companies already make about trust and access. The regional reports give those decisions more urgency without making every new request an attack.

A company is better placed to respond when permissions have owners, sensitive changes have an independent approval route and staff know they can raise a concern. Those expectations need to hold during a busy afternoon as well as during a policy discussion.

Mike Cheong helps Malaysian teams develop judgment around AI-assisted work, including reviewing outputs and recognising decisions that need human approval. Organisations can contact Mike Cheong to discuss workplace AI learning that fits their teams.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *