Why Awareness Training Only Works When It's Built Like a Control
I keep a running mental list of "things that will eventually bite every IT team," and social engineering has been at the top of it for years. Three reports landed in the same week in November 2026 and, taken together, they make an uncomfortable point most of us already suspect: you can patch every server, enforce MFA everywhere, and tune every email filter, and the incident still starts with someone making a normal-looking decision on a busy Tuesday.
Proofpoint's 2026 Singapore data says 92% of CISOs now call human risk their organisation's biggest cyber vulnerability — up from 61% a year earlier. Singapore's Cyber Security Agency (CSA) issued an advisory on 13 November warning that reported incidents have tripled, listing social engineering alongside exposed systems, weak identity controls, and supply-chain risk. And ENISA's third Cybersecurity Awareness Raising Conference, held in Stockholm on 11 November, was themed "From Awareness to Impact" — a pretty direct statement that attendance metrics are no longer the point.
This one's for the general IT professional: the sysadmin, the helpdesk lead, the security generalist who owns email, identity, and endpoints without a dedicated behaviour team. Here's what the numbers actually say, where the popular "training works" claim breaks down, and what a sensible programme looks like when you also have a day job.
Three Reports, One Week
| Source | Date | What it said |
|---|---|---|
| Proofpoint 2026 Voice of the CISO (Singapore) | 18 Nov 2026 | 92% of Singapore CISOs name human risk as their biggest cyber vulnerability (61% in 2025) |
| CSA Advisory AD-2026-017 | 13 Nov 2026 | Monthly reported incidents since September are 3× the January–August average; four common gaps, one of which is social engineering |
| ENISA 3rd Cybersecurity Awareness Raising Conference | 11 Nov 2026 | Theme: "From Awareness to Impact"; focus on measuring behaviour change, gamification, and AI |
Two wording cautions before we go further. First, Proofpoint asked about the biggest cyber vulnerability, not the "top concern" in the abstract. Second, CSA does not say humans are the number one cause — it lists four common factors, and unpatched, exposed systems come first. The honest reading is that behaviour is the common thread across most of those factors, not the only cause. Patching is a human process too. So is checking that a vendor's remote-access account still needs to exist.
What the Numbers Actually Show
Proofpoint's Singapore findings are more interesting than the headline statistic.
- Fewer incidents, worse concentration of risk. The share of CISOs expecting a material cyberattack in the next 12 months fell from 82% to 70%. Reported material data loss fell from 91% to 63%. Yet 64% still say their organisation is unprepared for a targeted attack.
- The insider angle is hard to ignore. Among organisations that did suffer material data loss, 44% cited malicious or criminal insiders, 43% compromised insiders, and 32% careless insiders. 97% said departing employees played a role.
- AI has made the human problem bigger, not smaller. GenAI security concerns jumped 30 points to 88%; 91% believe employees are likely to use AI in ways that expose sensitive data, and 84% are expected to manage AI risk without extra resources or expertise.
The survey covered 100 CISOs in Singapore, conducted by Censuswide in May 2026, as part of 1,600 CISOs across 16 markets at organisations with 1,000+ employees. That's a useful benchmark, not a census — but the direction of travel is hard to dismiss.
CSA's advisory grounds this in operational reality. It describes opportunistic scanning rather than a coordinated campaign against Singapore: attackers look for exposed servers, remote-access ports, and APIs without proper authentication, weak or reused passwords, missing MFA, vulnerable vendor-managed systems, and employees who will click a convincing phish. Its immediate measures are patching, auditing internet-facing assets, enforcing phishing-resistant MFA, and reviewing least privilege. Its longer-term list includes third-party oversight, continuous monitoring, tested backups and incident plans — and, explicitly, regular phishing simulations and staff training to counter AI-assisted social engineering.
Notice the shape of that list: technical controls and human controls are interleaved. CSA isn't offering training as a substitute for patching. Neither should you.
The Uncomfortable Part: Does Training Actually Work?
Here's where the easy article would end with "yes, train your staff." The evidence is genuinely mixed, and an IT professional deserves the real version.
The optimistic case. KnowBe4's 2026 Phishing by Industry Benchmarking Report reports a global "Phish-prone Percentage" (the share of employees who engage with a simulated phishing message) of 33.2% before training, dropping to 20.1% at 90 days, 4.2% at 12 months, and 3.9% at 24 months — an 87% reduction. Its 2025 edition analysed 67.7 million simulations across 14.5 million users. That's a large, consistent dataset.
The sceptical case. A 2025 randomised study from Purdue University put 12,511 employees at a fintech firm into lecture-based training, lecture-plus-interactive-exercises, and a control group. Result: click rates of 10.5% for trained staff versus 9.8% for the untrained (p=0.450), and reporting rates of 9.7% versus 8.9% (p=0.417). Statistically, no meaningful difference. A 2025 IEEE S&P study of 19,500 healthcare employees reached a similar conclusion, and a Computers & Security scoping review found near-term effects well documented but evidence for sustained behaviour change limited.
Why both can be true. The Purdue study found that what really moved click rates was the difficulty of the lure: 7% for easy messages, 8.7% for medium, and 15% for hard ones. It also found "inoculation" effects — in 36–55% of campaigns, one employee's report arrived before the first click, giving defenders a head start. Meanwhile, the vendor benchmarks measure customers who bought training and ran continuous simulations for a year, not a one-off course measured against a control group.
So "awareness training works" is too broad. A more accurate statement is:
Continuous, realistic, active practice with feedback can measurably reduce susceptibility. A once-a-year compliance module mostly produces a completion certificate.
If your programme is the latter, the academic sceptics are talking about you.
What "From Awareness to Impact" Looks Like in Practice
ENISA's Stockholm programme is useful precisely because it stopped asking "how many people attended?" and started asking "what changed?" Session one was titled Measuring Impact in Awareness Activities, with Prof. Steven Furnell on the Cyber Games Lab project, Joakim Kävrestad on "Measuring Impact — a continuous challenge," and an industry talk on nanolearning and teachable moments. The 2025 edition in Zagreb covered AI, measurement, and gamification — escape rooms, microlearning challenges, and tools like ENISA's AR-in-a-Box, which includes an online Cyber Awareness Game.
Translated for an IT team with limited time, the principles are:
- Short and repeated beats long and annual. Ten-minute modules and monthly simulations beat a 60-minute onboarding video nobody remembers. The KnowBe4 data suggests most of the gain happens between month three and month twelve — the first campaign isn't the programme.
- Make it active and specific. Show people an actual email from your own environment: the fake invoice, the shared-document invite, the "IT department" password reset. Ask them to spot the cue, not recite a policy. Active learning with detailed feedback is the component most reviews associate with better outcomes.
- Calibrate difficulty. If your simulations only use obvious lures with spelling mistakes, you'll hit a lovely 2% click rate and learn nothing. Mix in messages with few cues and strong premise alignment — the ones that look like your real business. The NIST Phish Scale is a free way to classify that difficulty.
- Make reporting effortless and blameless. A one-click "Report Phish" button, a helpdesk that thanks people instead of shaming them, and a feedback loop that feeds confirmed indicators back into your filters. That reporting-first loop is the "inoculation" effect the research found — and it's arguably more valuable than the training itself.
- Measure behaviour, not completions. Track repeat-click rate, time-to-first-report, and whether reported messages beat real clicks. Completion percentages are a compliance artefact.
- Keep the technical layer doing its job. Phishing-resistant MFA, DMARC/SPF/DKIM, conditional access, DLP on public AI tools, and prompt offboarding of dormant accounts. Proofpoint's data on departing employees and CSA's list of exposed systems are the same lesson from two angles: design for the assumption that someone will click.
That last point is the one worth pinning above the helpdesk. Awareness is a control, but it's a compensating control, not a replacement for identity and exposure management.
A 30/90-Day Runbook for IT Teams
You don't need a behaviour-change consultancy to start.
First 30 days
- Inventory internet-facing assets, remote-access ports, and APIs; disable what you don't recognise. This is CSA's number one gap and takes an afternoon with a spreadsheet and a shodan-style check.
- Enforce phishing-resistant MFA on privileged accounts, remote access, email, and cloud apps. Turn the "Report Phish" button on and tell everyone what it does.
- Audit dormant and shared accounts, especially contractors and leavers. The 97% departing-employee finding in Proofpoint's data makes this non-optional.
- Send one baseline simulation with realistic difficulty and record the result. Don't announce a target click rate.
By day 90
- Run short monthly simulations with varied difficulty; give immediate, specific feedback at the moment of failure.
- Publish the metric that matters: reports before first click, and repeat-click rate. Drop completion-rate dashboards from your security reports.
- Run a 20-minute tabletop or CTF-style exercise: hand a small team a suspicious email, a vendor access request, and an offboarding ticket, and have them talk through the response. ENISA's AR-in-a-Box and the Cyber Awareness Game are free starting points.
- Put guardrails around public GenAI use — guidance plus DLP and logging — rather than relying on a blanket block that people will route around.
- Re-check the baseline. Improvement over 90 days is a signal; a flat repeat-click rate means the programme is theatre.
For Singapore SMEs, CSA's SG Cyber Safe resources include role-specific toolkits and a CISO-as-a-Service scheme with up to 70% co-funding for eligible organisations. That's a far cheaper way to get help than learning these lessons through an incident.
Debrief
The headline is directionally right: humans are still the common thread in most breaches. But the useful conclusion isn't "blame the user" or "buy more training." It's this:
- Behaviour is one of several common gaps, and technical exposure still leads CSA's list.
- Continuous, realistic, measured practice can reduce risk; one-off awareness modules have weak evidence behind them.
- The fastest measurable win may be your reporting loop — one employee spotting a phish before anyone clicks — plus the technical controls that assume the click will happen anyway.
Treat people as part of the security architecture: give them good tooling, realistic practice, and a clear way to report. Then build the system so that a mistake is an event, not a breach.
Sources
- Proofpoint — 2026 Voice of the CISO, Singapore press release, 18 Nov 2026
- Proofpoint — 2026 Voice of the CISO, global press release, 9 Sep 2026
- CSA — Advisory on Recent Uptick in Reported Cyber Incidents, AD-2026-017, 13 Nov 2026
- ENISA — 3rd European Cybersecurity Awareness Raising Conference, "From Awareness to Impact", 11 Nov 2026
- ENISA — 2nd Cybersecurity Awareness Raising Conference, "Empowering the Human Element", 27 Nov 2025
- ENISA — Custom-made Awareness Raising to enhance Cybersecurity Culture (AR-in-a-Box), 10 Apr 2024
- KnowBe4 — 2026 Phishing by Industry Benchmarking Report findings, 7 Jul 2026
- Rozema & Davis (Purdue) — "Anti-Phishing Training (Still) Does Not Work", arXiv:2506.19899, Oct 2025