In the race to adopt artificial intelligence, Malaysian organisations have charged ahead with remarkable speed. A staggering 91% have embraced business AI tools, according to IBM's 2026 Cost of a Data Breach Report.
Yet this rapid adoption has created a dangerous blind spot. The same employees who paste source code, client lists and unreleased financials into chatbots to finish reports faster are becoming the weakest link in corporate security.
Abhishek Kumar Singh
As Abhishek Kumar Singh, head of security engineering in Singapore at Check Point Software Technologies, observes: "AI doesn't just introduce new security risks; it acts as an amplifier for all the architectural flaws we've already been sweeping under the rug."
The shadow AI surge
The numbers are sobering. IBM's research, based on 602 breached organisations globally, reveals that shadow AI incidents — the use of unapproved AI tools without organisational oversight — more than doubled in a single year, rising from 20% to 43% of breached organisations.
These incidents now carry an average price tag of $5.39 million. Even more troubling, 68% of breached organisations have no AI governance policy and no means to detect unauthorised AI use by employees.
This problem is not confined to distant markets. Malaysia's Personal Data Protection Department (JPDP) recently issued three landmark guidelines under the PDPA 2010, covering data protection impact assessments, data protection by design, and automated decision-making and profiling. The message is clear: accountability rests squarely with the organisation, not the tool.
According to Check Point's 2026 Annual AI Security Report, high-risk enterprise AI prompts have doubled over the past year, climbing from roughly 1 in 50 interactions to 1 in 25. While the average organisation runs ten AI applications monthly — many unapproved — between 87% and 93% experience at least one high-risk interaction each month.
Perhaps most revealing: most enterprise data exposure comes from ordinary, approved use, not from attacks, as employees share more context than they realise to get a useful answer.
The visibility gap
The challenge lies in the threat's mundane nature. An employee dropping sensitive data into a public AI prompt looks no different from ordinary web browsing — standard SSL/TLS encryption renders the traffic unreadable without specialised technology.
As Francis Yeoh, country director at SearchInform Malaysia, explains, traffic to AI services is encrypted with standard SSL/TLS, making it unreadable without MITM technology. Protecting users therefore requires adding a proxying layer to the usual firewall setup.
Francis Yeoh
"When someone drops sensitive data, e.g., a customer database, into a public AI prompt, they just want to finish with a business task faster — but the moment they hit paste, the data has left the building." Francis Yeoh
This paraphrases the practical risk Yeoh outlines: employees paste client lists or contracts into models to reformat or summarise them during business hours, and each action feels like ordinary work.
This is not hypothetical. Recent incidents show how easily AI tools can become vectors for data exposure. In one high-profile case, a Vercel employee connected an AI application to Google Workspace with full permissions, and when the tool was compromised, attackers gained access to internal dashboards, employee records, API keys and tokens — all without cracking a single password. The breach reportedly carried a $2 million ransom demand. No firewall was breached. No sophisticated exploit was used. An employee simply clicked "Allow All".
Singh reinforces this point, noting that modern AI tools "hook directly into our collaboration platforms, internal knowledge bases, cloud repositories, APIs, and business applications." Each integration opens a new pathway, and AI is "exceptionally good at finding, summarising, and redistributing information at a massive scale."
The risk compounds when organisations have existing structural issues: "If you already have structural issues, like sloppy data segmentation, sprawling service accounts, or over-permissive access rights, AI weaponizes those flaws instantly," said Singh.
A layered defence
Yeoh advocates for a practical, layered approach that balances security with operational needs. "The legislative answer is unambiguous: personal data must not be distributed this way," he states. His recommended three-pronged strategy combines technical controls, content-based filtering and employee training.
The technical layer begins with a Next-Gen Data Loss Prevention (DLP) system that monitors data transfers to AI services at both network and desktop application levels. Unlike AI Firewalls that only inspect web traffic, Next-Gen DLP operates at the agent level, embedding into desktop AI clients and covering popular messaging formats like WhatsApp and Telegram — the messengers through which most Malaysian business communication flows.
Content-based filtering offers three approaches, varying in flexibility. A blanket ban on non-corporate AI services is the easiest to deploy but least durable — employees inevitably find workarounds. A more effective approach is to block only transfers containing personal data, allowing harmless requests to pass through.
The most sophisticated approach involves streaming anonymisation, where identifiers are depersonalised before reaching the model. For example, "Ahmad bin Abdullah" becomes "AhXXd bin AbdXXXXh", while the AI still returns a usable result and regulated data never leaves the perimeter in identifiable form.
The third layer is training. Employees need a clear, applicable guideline: a short list of sanctioned services, a definition of what may never enter a prompt, and an understanding that once submitted, a prompt cannot be recalled.
"The two layers depend on each other," Yeoh warns. "Technical control without understanding breeds evasion, while training without control produces no record of whether the rules were actually followed — and that record is exactly what a regulator asks for."
Before deploying advanced technical controls, Singh argues that security teams must validate that fundamentals work.
Before giving the entire workforce keys to advanced AI tools, organisations should prioritise several core areas: tighten identity controls by auditing everything with access — human users, admins, third-party service accounts, and the AI integrations themselves; enforce least privilege so AI applications can only touch specific data and functions required for authorised workflows; map and classify data before plugging external LLM pipelines into repositories; upgrade telemetry to monitor, log and alert on AI-connected APIs and cloud services in real time; and manage configuration drift as environments change constantly.
"Point-in-time audits give you a nice snapshot of your risk posture on a Tuesday afternoon, whereas by Wednesday, a developer has spun up a new integration, and configuration drift has already set in," Singh cautions. "Building genuine resilience requires continuous validation." This insight aligns with Yeoh's emphasis on visibility — you cannot protect what you cannot see.
The regulatory stakes
For Malaysian organisations, the PDPA heightens the stakes considerably. Once regulated data enters an ungoverned external model, proving compliance becomes nearly impossible. The newly released PDPA guidelines make clear that AI cannot be the sole basis for decisions that significantly affect individuals, and that profiling always requires an impact assessment. The liability is unambiguous — and it rests with the organisation.
Take away
The narrative that data breaches come from sophisticated external attackers is dangerously outdated. In 2026, the greater threat often sits at the nearest desk, armed with good intentions and an AI chatbot.
Yeoh's practical focus is clear: inspect prompts and traffic for confidential content and block or anonymise regulated data before it leaves the enterprise. Singh echoes the broader urgency: "The winners in the AI era won't necessarily be the ones who cross the finish line first. Instead, they'll be the ones who manage to scale securely."
Malaysian organisations that fail to close the gap between AI adoption and governance will find themselves not only with multimillion-ringgit losses but also on the wrong side of regulators.
By locking down identity management, enforcing least-privilege principles, and shifting to continuous validation early on, organisations can stop playing catch-up and start driving AI transformation with real confidence.
The tools and strategies exist — the question is whether organisations will act before the data leaves the building.
