Monday 19 October 2026 · 8-minute read
Trust is the thing this week's stories keep circling. We extend it to software that logs in as us, to a voice on the phone that sounds exactly like the chief executive, and to almost any message that arrives looking plausible. None of that is a flaw in your colleagues; it is how human beings work, and it is also what this autumn's attackers, researchers, and regulators are converging on. The question running through everything below is not how to make people less trusting, but where to place the checks so that trust can survive contact with reality.
The newest insiders don't have a pulse
Trellix has surveyed 600 CISOs at organisations with more than 1,000 employees across the Americas, Europe, the Middle East and Asia-Pacific, and the headline lands hard: 95 per cent said an AI agent or other autonomous system was linked to an insider incident in the past twelve months. Handle that figure with care, since "linked to" stretches from cause to bystander, and Trellix sells tooling in exactly this space. The underlying shift still looks real. Among organisations reporting incidents, careless human action accounted for around 43 per cent of causes, deliberate misuse 29 per cent, and AI or autonomous systems 28 per cent. A full 78 per cent of respondents now count AI agents and machine identities inside their definition of an insider. At the same time, 63 per cent admitted they lack visibility into which AI tools employees actually use, 92 per cent said their personal accountability for insider incidents has grown this year, and 69 per cent worry about their own liability.
What does this mean for me?
Human risk programmes have always drawn their boundary around people, and that boundary is blurring. Every agent in your organisation acts on access a person delegates to it, prompted by a person, and supervised (or not) by a person. The behaviours worth working on are the human ones around the machine: who grants an agent access and on what basis, whether people disclose the tools they lean on, and whether anyone reviews what an assistant did with their credentials. A practical first step costs nothing: add one question to incident reviews and team check-ins, namely "what did automated tools do with your access this week?" The visibility gap CISOs report is at root a disclosure norm problem, and disclosure norms are squarely your territory.
Source: Trellix global CISO survey, reported by Help Net Security (14 October 2026). Vendor research, so read the superlatives accordingly.
Nobody wants to keep the boss waiting
KnowBe4 chose Singapore International Cyber Week to publish a Censuswide survey of 500 office workers with an uncomfortable finding at its centre. Yes, 36 per cent were not confident they could tell a genuine message from their CEO apart from one generated with AI. But the detection question is less interesting. Half of respondents said they would feel uncomfortable questioning a senior executive's request, 45 per cent feared delaying a senior leader's request as much as they feared causing a security incident, and 9 per cent said outright that they would rather risk a breach than keep the boss waiting. Then comes the inversion: senior managers were the most confident they could spot a fake (78 per cent), yet the most likely to comply with a cloned CEO voice before verifying it, at 35 per cent against 16 per cent of clerical staff. Only 26 per cent had received any training on AI-enabled scams that impersonate leaders.
What does this mean for me?
A cloned voice does not need to fool anyone's ears. It needs people to feel unable to say "can I just check?", and that is a culture problem wearing a technology costume. The highest-value interventions here are social, not technical. Get a senior leader to state, publicly and more than once, that verification is a mark of professionalism and that they expect to be checked. Rehearse the mechanics until they are boring: any urgent request involving money, credentials or data gets a callback on a number you already hold. And measure the thing that matters, because a single culture-survey item ("I would feel comfortable double-checking an unusual request from a senior leader") will tell you more about your deepfake resilience than any amount of detection training. It is one vendor's survey of one country, but the hierarchy effect it describes will travel.
Source: KnowBe4, "Would You Know Your CEO's Real Voice?", reported by Bernama via Malay Mail (14 October 2026).
Believing is the default setting
The October issue of the American Psychological Association's Monitor on Psychology carries a long feature on why fraud keeps working, drawing on researchers including Stacey Wood (Scripps College), Yaniv Hanoch (University of Wolverhampton) and Marti DeLiema (University of Minnesota). The through-line is truth bias: humans default to assuming messages are credible, because life would be unliveable otherwise, and scammers simply rent that default by adding urgency, emotion and personal detail scraped from social media. The research punctures the comforting idea of a victim "type". Circumstances such as loneliness, financial strain, overconfidence and prior victimisation predict vulnerability far better than age or intelligence, and AI polish has quietly retired the spelling-mistake heuristic many of us still teach. The detail worth stealing: in one study, truth bias appeared when people were forced to label statements true or false, and vanished when they were allowed to answer "unsure". The piece is also clear-eyed about the aftermath, noting Global Anti-Scam Alliance data suggesting only around 4 per cent of victims recover their money, while many absorb blame on top of the loss.
What does this mean for me?
Three concrete moves. First, audit your training content for typo-spotting advice and retire it; it now actively miscalibrates people. Second, make "I'm not sure" a legitimate, celebrated state: reporting routes that offer a low-friction "this might be nothing" option will catch what binary report-or-ignore flows miss. Third, look at how your organisation treats employees who are successfully phished. The research on victim shame maps directly onto incident reporting: people who expect blame hide incidents, and hidden incidents are the expensive kind. This journalistic feature synthesises peer-reviewed work rather than being a study itself, but it is an unusually good one-stop summary of the field.
Source: Monitor on Psychology, APA, October 2026 issue.
Who pays when the customer presses send?
Away from the awareness industry entirely, AMRO (the ASEAN+3 Macroeconomic Research Office, the body that monitors financial stability for Southeast Asia plus China, Japan and Korea) has published an analytical note on digital scam risk, co-authored with the Bank of Thailand. The scale it documents is sobering: roughly 55 per cent of adults across the surveyed economies faced at least one scam attempt in 2024 to 2025, with survey-based loss estimates ranging from 0.2 to 3.4 per cent of GDP depending on the economy. The recovery gap is starker still, with only 20 per cent of victims in the ASEAN economies recovering any funds, against 46 per cent in China, Japan and Korea. The policy direction is the real story for our field. Because victims authorise these payments themselves, the deception is complete before any bank control fires, so regulators are moving liability and friction into the system: Singapore's Shared Responsibility Framework and Thailand's 2025 royal decree assign duties across banks, telecoms firms and platforms; Hong Kong is consulting on a model that weighs customer circumstances, and Thailand reports 3.4 million mule accounts suspended, with authorised fraud losses falling markedly since 2024.
What does this mean for me?
The centre of gravity in counter-scam policy is shifting from "educate the user" to "design the system and share the loss", and Asia is currently writing the playbook that European and North American regulators will be reading. If you work in financial services, these frameworks are worth tracking directly. For everyone else, the behavioural lesson transfers neatly inside the organisation: Thailand's gains came from friction and checks at the moment of payment, not from warnings issued weeks earlier. Your invoice approval and payment-change processes deserve the same logic, with the pause built into the workflow rather than left to the vigilance of whoever is busiest that day. One caveat: this is a working-style note reflecting its authors' views, and the loss estimates are survey-based with wide ranges.
Source: AMRO analytical note, Managing Digital Scam Risks in ASEAN+3 (12 October 2026).
Until next week
Four stories, one thread: trust is not the vulnerability; unexamined trust is. If a single idea deserves to travel from this edition into your programme, make it the small one from the psychology research: give people a respectable way to say they are not sure yet. See you next Monday.