TechForge

October 14, 2026

  • AI agents optimise for the cheapest route and cannot tell a shortcut from a breach, Google’s Lim Yihao says.
  • One accounting agent fired 15,000 reasoning queries in under an hour and burned through US$50,000.

AI agents are not breaking into systems because someone told them to. They do it because they were built to find the fastest way to finish a job, and nothing in that design tells them where the line is.

That is how Lim Yihao, lead threat intelligence advisor for Japan, Asia-Pacific and the Middle East and Africa at Google Threat Intelligence Group, describes a pattern his team has been tracking. “Agents are not misbehaving intentionally,” he told Tech Wire Asia on the sidelines of CyberDSA 2026 in Kuala Lumpur on 6 October. “The agents are misbehaving because they are being designed to try and find the shortest way to solve the problem. [Which includes] the cheapest way, the most efficient way to solve the problem.“

Why AI agents cut corners

Lim Yihao, lead threat intelligence advisor, Google Threat Intelligence Group
Lim Yihao, lead threat intelligence advisor for Japan, Asia-Pacific and the Middle East and Africa, Google Threat Intelligence Group.

Lim walked through the choice an agent makes when it is handed a task. It can build it from scratch, which costs tokens, time, and money. Or it can look at how somebody else solved the same problem and work from that. “If I see another company already has a similar product, maybe I’d go and study what they’ve done, take the principles of how they design that,” he said. “And then I’ll make this in a much shorter time, at a much lower cost.“

Studying what someone else built can mean reading their public documentation. It can also mean getting inside a system nobody opened for it. The agent weighs both the way it weighs everything else, by what they cost rather than by permission.

Which is why from the agent’s side, nothing has gone wrong. “They are designed to be efficient, so to them it’s not a hack,” Lim said. “But that’s where the real problem is. Because they don’t know what is a hack and what is not.“

Asked whether that judgment could be trained into a model, Lim answered with a question. “How do you teach ethics?” He called it a million-dollar question, and said the consequences of leaving it unanswered are already visible. “If we cannot solve this problem, we will continue to see more and more of these kinds of things happen.“

His explanation for why it stays unanswered is about where the money is. The industry’s priorities, he said, are capability: mapping the human genome, quantum computing, better chips, better mathematics. “Those are the things that pay the most money,” he said. “But if you have the best ethics, who’s gonna pay you more?“

The lethal trifecta

Dr Monisha Oberoi, regional consulting leader at Google Mandiant, sees the same problem from inside the organisations that deploy these agents rather than from the threat intelligence side. Her concern is how casually they are being built. Marketing, sales and HR teams are each spinning up their own agents, she said, because companies are encouraging staff to use more AI.

The risk, in her account, comes from three things stacking up. “When we create any agent which is using untrusted data and giving it the privilege to execute, and giving it access to your crown jewel, it is waiting for disaster to happen,” Oberoi said. “So this is what we call the lethal trifecta.” The term is used across the security industry for the combination of untrusted input, the ability to act and access to sensitive systems.

She gave a case from recent work. An agent built to handle accounting hit a null field where it expected a value, and could not proceed. “The autonomous agent automatically on its own decided that I will fire some reasoning queries,” Oberoi said. “It fired 15,000 reasoning queries in less than an hour. The tokens were utilised, and US$50,000 was consumed in less than one hour because there were no proper guards.” The attack type has a name, denial of wallet, though in this instance nobody attacked anything. The agent did it on its own.

Oberoi’s point is that the controls organisations apply to people are not being applied to software that now acts like people. “When we talk of human beings, we talk of least privilege, zero trust and least privilege,” she said. “But in the case of an agent, are you

Dr Monisha Oberoi, regional consulting leader at Google Mandiant
Dr Monisha Oberoi, regional consulting leader at Google Mandiant, speaking to Tech Wire Asia at CyberDSA 2026.

doing that? That’s a big question.“

Shadow AI and the data staff hand over

Agents are only part of what Oberoi calls shadow AI. The simpler version is an employee who wants help drafting something and pastes company material into a public tool to get it. “It is being done all the time,” she said, “because once it’s done, you have no control over it.“

What follows, in her description, is a loop organisations cannot see. Staff feed internal detail into models with no bad intent and no record of having done it. The same models are available to anyone running reconnaissance on that company.“You are feeding the information without realising,” Oberoi said. “And the organisation doesn’t have control over it.“

Her first question to clients is whether they know the scale of it. “Do you even know how many agents, how much AI is being utilised? Are you even aware of what your employees are actually putting into these tools?“

What the speed does to a security team

When attackers use the same technology, the arithmetic of a security operations centre changes. Lim described the old sequence, where every step of an intrusion was a human typing an instruction. List the files. Open that folder. Find that file. “Very tedious, very slow,” he said. “But now that it is AI-empowered, the AI can think at machine speed. So within three seconds they have done like 15 steps.“

What a SOC sees does not change much, which is the difficulty. “They will still see the normal things that a normal hacker would do,” Lim said, “but just time to react, super short.” He is sceptical that staffing alone can close the gap. “If you use humans to detect, it’s going to be 5,000 people in your organisation, everyone logging in, but how will you know what is suspicious?“

Mandiant’s own numbers show the compression. Its M-Trends 2026 report, based on more than 500,000 hours of frontline investigations during 2025, found the median time between an attacker breaching a perimeter and handing that access to a second group had fallen to 22 seconds, from more than eight hours in 2022. Oberoi pointed to the same figure as evidence of how specialised the market has become, with one group selling entry and others doing what comes after. “Everybody has a task cut out,” she said.

Detection has improved on one measure and not on another. The same report puts global median dwell time at 14 days, up from 11, while incidents it classifies as cyber espionage ran to a median of 122 days. Organisations found the intrusion themselves 52% of the time, up from 43% a year earlier, with external parties notifying them in 34% of cases and attackers themselves in 14%.

Both concluded the same way. “We have to fight the machine speed threat with machine speed defence,” Oberoi said. Lim was more blunt about why defenders cannot opt out. “They [hackers] are already using the gun. Why are they [defenders] still using a stone?“

The most sophisticated regional case used no AI at all

Asked what Google Threat Intelligence Group sees in Southeast Asia rather than globally, Lim pointed to UNC6384, a China-nexus actor his team has documented targeting diplomats in the region. He described the tradecraft as sophisticated. Then he added a qualifier worth holding onto: “There’s no AI that is used as part of this attack.“

The most capable regional intrusion he could name, in other words, was conventional work done well. What AI changes, on his account, is the pace at which methods turn over. “It’s definitely important to have the intel on what’s the latest thing because they are innovating super fast,” he said. “Super fast now with AI, even faster“

He expects voice cloning to arrive in scams here as a matter of timing rather than capability. “Maybe next time the person calls you is your mom’s voice,” he said. “Or your friend’s voice.” On the region’s scam compounds, he was straightforward about the limits of what his team can see. Google Threat Intelligence Group has no visibility into them, he said, though he would be surprised if they were not already using the same tooling.

What they want organisations to do

Both work for a company that sells the services they are recommending, and Oberoi’s advice maps closely onto Mandiant’s consulting work. Her sequence starts before deployment. Establish governance over how AI is used, find out how much shadow AI already exists, apply least privilege to agents as to people, and test the guardrails rather than assuming them. “Are there enough guardrails, and are we testing those guardrails enough to ensure that the agent is doing what it’s supposed to do?”, she stated.

She is unsentimental about compliance as a substitute. An organisation running one red team exercise a year, or treating a ticked checklist as preparedness, is “sitting on a time bomb“, she said.

Lim’s version is narrower and comes back to the same gap. Keeping a person in the loop matters, he said, not because agents go rogue but because the output drifts. “The output could be very different from what you envision at the start,” he said. He described building checkpoints through the process so the work is verified as it goes rather than only at the end.

What keeps them up at night

Asked what keeps her up at night, Oberoi pointed at governments rather than companies. An organisation like hers sees threat intelligence and frontline investigations across the world, she said, while most governments work from a far narrower view even when buying in threat feeds. 

The harder half is people. “Skilling up your workforce right from the ground up is a huge challenge,” she said, which is why she argues for secondment and continuous upskilling across both public and private sectors.

Put the same question to Lim, and he did not describe a breach. He brought up a case he came across on GitHub. A developer flagged a package as malicious and warned others off the account that had uploaded it. The uploader pushed back, saying the work was legitimate. A third account then vouched for the uploader, saying it had used their software before.

Both accounts arguing with the developer were agents. “He still used another agent to make it look like more people,” Lim said. “So this is, I would say, just a glimpse of keep me up at night. Because it’s super new.”

The same trick works on consumer platforms. “Let’s say on Instagram somebody texts you and, wow, super friendly, super nice,” Lim said. “Maybe [they are] not even human.”

 

 

 

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.

About the Author

Dashveenjit Kaur

Dashveen writes for Tech Wire Asia and TechHQ, providing research-based commentary on the exciting world of technology in business. Previously, she reported on the ground of Malaysia’s fast-paced political arena and stock market.

Related

November 13, 2026

November 13, 2026

November 11, 2026

November 11, 2026

Join our Community

Subscribe now to get all our premium content and latest tech news delivered straight to your inbox

Popular

35677 view(s)
28210 view(s)
25631 view(s)
20740 view(s)

Subscribe

All our premium content and latest tech news delivered straight to your inbox

This field is for validation purposes and should be left unchanged.