MOST of the AI and cybersecurity conversations over the past two years have focused on AI
writing better phishing emails or cloning a voice for a fraud call. That’s real, but it understates
what’s happening now. What’s different about agentic AI is autonomy. An attacker can hand an
AI agent a goal rather than a script, and the agent will scan for exposed services, adapt when a
defense blocks it, and chain together multiple stages of an intrusion with very little human input
in between.
A threat intelligence disclosure in June 2026 gave the industry a concrete look at this, when
researchers uncovered a fully operational threat-actor lab built specifically to develop malware
capable of defeating endpoint detection tools. The setup used an AI system as a coordinating
“manager” directing several agents through code generation, testing, and revision, running each
build against live security products before shipping it. It wasn’t theoretical red-teaming, it was a
working pipeline, industrializing what used to be slow, manual malware development.
The bigger structural story sits in Sophos’ most recent Active Adversary Report – composed
from 661 incident response and MDR cases. Here we found that median attacker dwell time has
decreased to three days, with identity-based attacks, stolen credentials, weak MFA, poorly
protected accounts, now the dominant ways attackers get in, ahead of software exploitation.
This is reinforced by Sophos’ State of Identity Security 2026, which found that 71% of
organisations suffered at least one identity-related breach in the past year, while 67% of
surveyed ransomware victims said their ransomware incident stemmed from an identity attack.
AI isn’t inventing new ways to get through the door. Instead, it’s collapsing the time between
getting in and doing damage, which is arguably the more dangerous problem for defenders
operating on human-paced response cycles. Malaysia’s own numbers point the same way.
CyberSecurity Malaysia’s Cyber999 centre logged a 29% jump in reported data breach
incidents in the first quarter of 2025 alone, while credential-focused threats continued to feature
prominently in reported incidents.
Why this is a governance problem, not just a tooling one
Malaysia’s regulators are asking similar questions. The Cyber Security Act 2024 and Bank
Negara Malaysia’s revised Risk Management in Technology (RMiT) policy have pushed
stronger governance expectations around incident reporting, risk assessment and cybersecurity
controls, while the Personal Data Protection (Amendment) Act 2024 adds another layer, with
mandatory data breach notification in force since June 2025 and and increased penalties for
breaching data protection principles.
The policy landscape is still evolving. In July, Parliament passed the Cybercrimes Bill 2026,
which is intended to replace the Computer Crimes Act 1997 and strengthen Malaysia’s
framework for increasingly complex cyber threats, including the misuse of technologies such as
AI. More recently, the Digital Ministry completed its draft Artificial Intelligence Governance Bill,
which is expected to be tabled in Parliament this quarter or by the first quarter of 2027 at the
latest. Together, these developments signal that AI and cybersecurity are increasingly becoming
governance and accountability issues, not simply technology decisions left to security teams.
Frameworks alone don’t close the human gap, though. In January 2026, BNM fined Bank
Kerjasama Rakyat Malaysia Bhd RM1 million after an external threat actor breached its IT
infrastructure, citing inadequate cybersecurity controls and incident response, not a missing
policy.
The challenge is that most governance frameworks were built for human-paced threats. Agentic
AI changes that equation by allowing attacks to move, adapt and escalate at machine speed,
forcing organisations to rethink not just their security tools, but how quickly decisions,
accountability and response can happen.
So what this means in practice?
None of this calls for panic, and it doesn’t mean the fundamentals have changed. It means the
fundamentals matter more. A few things worth prioritizing:
Identity first. With identity compromise now the leading entry point across incident
response caseloads, phishing-resistant MFA and tight management of both human and
machine (“non-human”) identities deserve to sit above most other line items on a 2026
security roadmap, particularly with the PDPA’s breach-notification duties and RMiT’s
identity controls both now in force.
Assume compressed timelines. Three-day median dwell time means detection,
containment, and response can no longer be planned as sequential steps with review
gates in between. They need to run in parallel, which usually means 24/7 monitoring
through a managed detection and response service or an equivalent always-on
capability, not a business-hours security operations center.
Match machine speed with machine speed. Signature-based detection was designed
for a world where new threats appeared occasionally. When an AI-assisted pipeline can
generate and test dozens of evasion variants in an afternoon, detection needs to work
off attacker techniques and behaviors, not just known signatures because the specific
payload may genuinely be new every time.
Close the human gap. Investment in tools without investment in people and forensic
readiness leaves the weakness that BNM’s RM1 million fine on Bank Rakyat put a price
on. Tabletop exercises and incident response drills matter more, not less, as AI
compresses response windows.
Agentic AI hasn’t replaced the attacker’s human intent, and human error on the defending side
still drives many incidents. What it has done is remove the friction that slowed attacks down and
gave defenders a little breathing room. That breathing room now needs to be re-engineered
back into enterprise defence through stronger identity controls, always-on monitoring and
detection built around attacker techniques rather than signatures, before it disappears for good. – November 2, 2026
***Malis Selamat is the Managing Director (Asean) of cybersecurity firm Sophos