2 November 2026 · 9-minute read
Con artists used to invest in a good suit for a reason. Most of us decide who to trust in seconds, from surface cues, and this week's research shows attackers putting serious effort into those surfaces: fake IT support arriving through your own collaboration tools, scam emails that pass every technical check we taught people to rely on, and AI systems quietly mass-producing the costumes. The thread running through all five stories is that looking legitimate has never been cheaper, so the cues we give employees need an urgent rethink.
The help desk that breaks things first
Arctic Wolf has documented more than 400 incidents in which attackers abused Microsoft Teams external chat to get inside organisations, and the choreography is worth studying. First, the attacker floods an employee's inbox with hundreds of newsletter and subscription confirmations. Minutes later, a friendly "IT support" contact appears on Teams, offering to fix the very problem they created. Display names do the costume work: "IT Service Desk (Matthew)", shield emoji, even "(Internal)" appended to names that Teams itself labels as external. Between 30 and 40 per cent of cases used free onmicrosoft.com tenants, which need no verification and can be set up in minutes. Once the victim engages, the fake technician moves to a voice call and talks them into running Quick Assist or another remote access tool. Malicious code can run in under 15 minutes, and some intrusions end in data theft and ransomware. The technique began with the Black Basta ransomware group in 2024 and has since spread to Cactus, 3 AM and Chaos operations.
The psychology is elegant and nasty. Employees have been trained for years to treat email with suspicion, so the attack simply moves to a channel that feels internal, and the manufactured inbox crisis makes the rescuer welcome rather than suspect.
What does this mean for me? Three things. Ask your IT team this week whether Teams external access is still on its default open setting, because this attack needs no vulnerability. Give employees one memorable promise to anchor on, something like "IT will never contact you out of the blue on Teams to fix an email problem", since a clear positive rule beats a list of warning signs. And teach the pattern, not the platform: a sudden flood of junk followed by an unsolicited offer of help should itself be the red flag, wherever the helper turns up.
Source: Arctic Wolf threat research, reported by Help Net Security (27 October 2026): Hackers bury inboxes in spam then show up on Microsoft Teams posing as IT support
When passing the checks means nothing
KnowBe4's eighth Phishing Threat Trends report lands on a similar theme from the email side: attackers have largely stopped faking legitimate infrastructure and started borrowing it. By the vendor's telemetry, 41.3 per cent of phishing emails now get past secure email gateways, peaking at 48.3 per cent in July, and of the messages the gateways miss, 58.1 per cent still pass DMARC, the authentication standard organisations spent years rolling out. In the third quarter, attacks were twice as likely to originate from genuinely legitimate platforms as in the quarter before. Phishing volume overall rose 29.8 per cent year on year, spear phishing rose 41.1 per cent, and impersonation of AI brands rose 228 per cent between April and August.
The usual caveat applies with extra force here: this is a vendor's view from its own detection ecosystem, and KnowBe4 sells the remedy it recommends. But the direction of travel matches what Arctic Wolf is seeing on Teams, and the implication for awareness content is hard to dodge.
What does this mean for me? Audit your own training materials for advice that quietly depends on infrastructure cues: check the domain, look for the padlock, hover over the link. When a majority of filter-evading phish passes DMARC, those cues no longer separate safe from unsafe, and teaching them may actively mislead. Shift the emphasis to behavioural signals that attackers cannot borrow: the unexpected request, the pressure to act now, the change of channel or payment details. If your simulations still rely on lookalike domains and spelling errors, they are rehearsing yesterday's attack.
Source: KnowBe4 Phishing Threat Trends report, vol. 8 (29 October 2026, vendor telemetry): When Bad Actors Use Reality Against Us
The riskiest AI user is on the payroll
While attackers borrow trust, employees are quietly borrowing tools. O'Reilly's new survey of information security professionals, conducted in September, found that the AI risks keeping practitioners awake are not AI-powered attacks but their own colleagues: overreliance on AI output without human verification and sensitive data leaking through AI tools were each named by 49 per cent of respondents, and 59 per cent believe teammates use AI outside official policy, with 23 per cent describing that use as widespread.
The finding worth pinning to the wall, though, is the policy effect. In organisations with a comprehensive AI security policy, only 17 per cent reported widespread shadow AI, against 35 to 37 per cent where policies were basic or absent. Only 44 per cent of organisations have such a policy. The release does not state the sample size, and perceptions of colleagues' behaviour are not measurements of it, so take the numbers with a grain of salt. The pattern, however, echoes decades of shadow IT research: people route around ambiguity, not around rules.
What does this mean for me? Treat shadow AI as a design problem rather than a discipline problem. A short, concrete policy that names approved tools and spells out what may and may not be pasted into them appears to do real work, so if yours is still a holding statement, this is the week to fix it. Pair it with an amnesty question in your next survey or champions call: which AI tools are you actually using, and for what? The answers will tell you where the safe path needs paving.
Source: O'Reilly press release (21 October 2026, industry survey): O'Reilly Survey Reveals Wide Gap Between AI Security Threats and Organizational Preparedness
The robots are not ready to protect grandma
A peer-reviewed, open-access review in the journal Information, by Balint Voros and Cecilia Sik-Lanyi at Hungary's University of Pannonia, asks a question most of our field skips: do AI-based scam defences actually work for older adults? The backdrop is grim. Americans aged 60 and over reported 7.7 billion dollars in losses to the FBI's Internet Crime Complaint Center in 2025, 59 per cent more than the year before, with voice cloning and deepfakes now woven into the scams. Reviewing 63 studies from 2019 to 2026, the authors found that most interventions for older users measure knowledge, confidence or intentions rather than real behaviour, that results were mixed or null where detection performance was actually measured, and that not one AI scam-detection system had been evaluated with older users in a field setting. One detail stings: the only study testing audio deepfake detectors on older speakers found the detectors roughly doubled their error rate.
What does this mean for me? Two lessons, one generous and one uncomfortable. The generous one: your employees are also sons, daughters and grandchildren, and "help protect your family from voice-clone scams" is both genuinely needed and some of the most engaging awareness content you can run, because the technology meant to catch these scams is not yet holding the line for this group. The uncomfortable one: the measurement gap the authors criticise, confidence scores standing in for behaviour, describes a fair share of corporate awareness programmes too. Worth a quiet look at your own metrics.
Source: Voros and Sik-Lanyi, University of Pannonia, peer-reviewed review in Information (25 October 2026, open access): Artificial Intelligence for Elder-Scam Defense
AI phishing sites: a production line, not a master forger
Finally, some useful nuance for anyone briefing leadership on AI-enabled threats. A research team led by Fujiao Ji and Doowon Kim released PhishGen, a preprint systematically testing whether large language models can build working phishing websites. Across 14 models, the answer is yes, but with texture. Some 96.2 per cent of outputs produced after bypassing the models' safety guardrails rendered as web pages, yet only 38.7 per cent completed the full credential-stealing pipeline, and human raters scored their resemblance to real sites a modest 2.02 out of 5. Two findings cut against the received wisdom. Mainstream commercial models were by far the most capable, with one reaching up to 94 per cent functional success while "uncensored" models rarely passed 45 per cent. And the strongest models outperformed the 64 real-world phishing kits the team benchmarked against. This is a preprint, not yet peer-reviewed, but the methods are unusually thorough.
What does this mean for me? The threat AI brings to phishing sites is scale and low barriers, not perfection. Expect more fakes rather than flawless ones, and note that visual inspection was already a dying defence given what the KnowBe4 data says about borrowed infrastructure. The durable advice is procedural: reach login pages from your own bookmarks, lean on password managers that refuse to autofill the wrong domain, and keep pushing passkeys, which do not care how convincing the page looks.
Source: Ji et al., research preprint on arXiv (21 October 2026, not yet peer-reviewed): PhishGen: Understanding LLM Capabilities in Automated Phishing Website Generation
Signing off
A week to retire old cues and teach better ones. If you do one thing before Friday, make it the Teams external access conversation with your IT team, and if you do two, ask what your metrics would show if confidence stopped counting as evidence. See you next Monday.