SANS 2026 Report: AI Is Now the #2 Human Cyber Risk
SANS Institute’s 11th annual Security Awareness & Culture Report, based on over 1,700 practitioners worldwide, shows AI has jumped from the 4th to the 2nd biggest human risk organisations face, behind only social engineering. Here’s the short version.
Report at a glance
Top 4 human risks for 2026 (AI jumped from #4 to #2):
- Social Engineering (phishing/vishing/smishing/deepfakes) – 77%
- Inappropriate AI Use at Work – 42%
- Sensitive Data Mishandling – 39%
- Weak Passwords/Authentication – 22%
Program maturity (bell curve, shifting toward more mature): 3% non-existent, 19% compliance-focused, 45% promoting awareness/behavior change, 21% long-term culture change, 12% optimization & resilience.
AI moved up fast because adoption outran policy
Two years ago AI ranked 4th on the list of top human risks. This year it’s 2nd, at 42%, just behind social engineering (77%). SANS is clear the issue isn’t that AI is inherently unsafe, it’s that organisations haven’t kept pace with policy and controls as employees adopted it. Three distinct risk areas now need separate attention: everyday GenAI use (shadow AI tools, oversharing sensitive data, blind trust in AI output), vibe coding (non-developers shipping AI-written code with no security review), and agentic AI (bots taking action with no human in the loop).
Most programmes are stuck in the middle
Using SANS’s five-stage maturity model, 45% of organisations sit at “Promoting Awareness & Behaviour Change” and 19% are still purely “Compliance-Focused.” Only 12% have reached the top tier, “Optimization & Resilience.” The gap between the middle and the top isn’t about content, it’s about consistency and reinforcement over years, not a once-a-year training push.
Team size and time are what actually move the needle
Six years of data now show the same pattern: it takes roughly 3 dedicated FTEs to shift behaviour, and 4.3+ to embed lasting culture change, sustained over 3-10 years. The top reported barrier, for the fifth year running, isn’t budget. It’s time.
What to actually do
- Treat AI as three separate risks – GenAI use, vibe coding, and agentic AI each need their own policy and training, not one generic AI module.
- Extend existing data policies to AI tools – don’t build new rules from scratch; apply the authorised-systems and authorised-data thinking you already use for cloud.
- Ditch the annual compliance module – shorter, continuous, role-specific reinforcement consistently outperforms a once-a-year training event.
- Replace “gotcha” phishing tests with positive reinforcement – organisations that publicly recognised good reporting behaviour, instead of just flagging failures, saw report rates rise up to 40%.
None of these controls work as a standalone fix, policy, training cadence, and positive reinforcement each cover a different part of the human risk surface. Used together, they’re what separates the 12% of organisations at the top of SANS’s maturity model from everyone else.
