Last updated: November 18, 2026
What makes AI security awareness training work? Completion rates do not predict it. A 2025 University of California San Diego study of more than 19,500 employees found no significant link between recent training and phishing failures. Effective AI security awareness training targets a few risky behaviors, such as pasting data into unapproved tools, and backs them with technical controls.
TL;DR: Key Takeaways
- A UC San Diego Health experiment with more than 19,500 employees found no significant relationship between recently completed annual training and falling for phishing.
- The same study found embedded training cut failures by only about 2%, while interactive training cut clicking by 19% in relative terms.
- Shadow AI use rose from 15% to 45% of employees in one year, according to Verizon's 2026 DBIR.
- The FBI's 2025 report recorded 22,364 AI-related complaints and $893 million in losses.
- NIST SP 800-50 Rev. 1 treats training as a life cycle that includes measuring program impact.
AI security awareness training at a glance
| Approach | What it changes | What the evidence says | Best used for |
|---|---|---|---|
| Annual course | Knowledge of policy | No significant link to phishing outcomes in the UCSD study | Compliance records |
| Embedded phishing training | Behavior at the moment of failure | About 2% fewer failures; most users spend a minute or less | Low-cost reinforcement |
| Interactive, scenario-based training | Specific decisions | 19% relative reduction in clicking in the same study | High-risk roles and behaviors |
| Technical controls | What is possible | Recommended by the study authors | Everything training cannot reliably fix |
Why most AI security awareness training does not change behavior
Most programs measure attendance, not decisions. The UCSD researchers ran an eight-month randomized experiment with 10 phishing campaigns and found that annual training did not predict who clicked. Most employees barely engaged with the lessons that followed a failure, which is consistent with the weak effect.
| Finding (UC San Diego Health, 2025) | Result |
|---|---|
| Employees in the experiment | More than 19,500 |
| Users who spent one minute or less on embedded material | 75% |
| Users who closed the page immediately | About one third |
| Users who completed embedded training | Under 24% |
| Employees who failed at least once | 56% |
| Failure rate across lures | 1.82% (password update) to 30.8% (vacation policy) |
The last row matters most for design. According to the researchers (2025), the lure mattered enormously: a routine password notice fooled few employees, while a vacation policy fooled almost a third. Training that treats employees as the problem misses that attackers choose the bait.
The authors recommend technical countermeasures such as hardware or app-based two-factor authentication and password managers that only fill credentials on the correct domain. Training still matters, but it cannot be the only layer.
What employees need to know about AI risks
Keep the curriculum short and tied to four situations employees meet daily.
Shadow AI. According to Verizon's 2026 DBIR, shadow AI use rose from 15% to 45% of employees in a year, and it ranked third among non-malicious data-leakage activities. Employees need a clear rule on which AI tools are approved and what data may never go into a prompt.
AI-enabled fraud. According to the FBI (2026), 22,364 complaints in 2025 referenced AI, with $893 million in losses. The FBI notes that AI is tagged only when the filer mentions it, so the true count is likely higher. Teach a call-back habit for any urgent request, whether it arrives by email, voice or video.
Hidden instructions in files. According to OWASP (2026), prompt injection remains the top LLM risk and now covers attacks hidden in images or audio. Employees who let an assistant summarize an untrusted document should know it can carry instructions.
Oversharing. Sensitive Information Disclosure ranks second in the same list. Show employees which data classes stay out of external models.
How to design AI security awareness training that changes behavior
Design the program around decisions, not modules. The method below follows NIST's life-cycle framing and the UCSD findings.
How we assessed this
We compared the UCSD randomized experiment, NIST SP 800-50 Rev. 1, Verizon's 2026 DBIR and the FBI's 2025 complaint data, and kept only claims that a named source supports. We did not use vendor benchmarks.
1. Pick three to five behaviors
List the actions that cause the most exposure, such as pasting customer data into a chatbot, approving an urgent payment by video call, or installing a browser extension. NIST's guidance asks programs to be employee-focused and tied to risk, so a short list beats a full catalog.
2. Train at the moment of the decision
Interactive training outperformed embedded training in the UCSD study. Place a short prompt where the choice happens, for example when an employee opens an AI tool, and keep it under a minute.
3. Give people a safe alternative
A ban pushes use out of sight. Offer an approved assistant, publish what it may handle, and make it easier to use than the unapproved one.
4. Make reporting rewarding
Count a report as a success, even when the message was harmless. Reporting rate tells you more about culture than click rate does.
5. Measure behavior, then adjust
NIST asks programs to measure impact. Track the metrics below quarterly and retire content that does not move them.
Metrics that show behavior change
| Metric | What it shows | Caution |
|---|---|---|
| Reporting rate | Whether employees raise suspicious items | Needs a simple report button |
| Time to report | How fast the response starts | Track the median, not the average |
| Failure rate per lure | Which scenarios fool people | Vary lures; difficulty differs widely |
| Unapproved AI tool use | Shadow AI exposure | Measured from gateway or proxy logs |
| Sensitive data blocked | Control coverage | Count blocked events, not people |
Avoid setting a target click rate from an industry average. The UCSD lure spread, from 1.82% to 30.8%, shows that one number depends mostly on the message you send.
AI security and governance controls that back up training
Training reduces mistakes, and controls limit what a mistake can expose. OWASP's 2026 list places Prompt Injection first and Sensitive Information Disclosure second, which are the two risks employees influence most.
Agent Gateway (TrustGate) gives security teams one governed route to model providers, so they can see which tools and models employees use, apply policy to prompts and log requests for audit. Agent Runtime Security (TrustGuard) adds checks when AI agents act on behalf of employees, which matters once assistants read untrusted files. NeuralTrust, a Barcelona-based AI agent security company, is recognized in six Gartner Hype Cycles for 2026, including AI Runtime Defense. See the AI Gateway and agent runtime security.
Which AI security awareness training approach should you choose?
| Your situation | Start with | Add next |
|---|---|---|
| No AI policy yet | An approved-tools list and a one-page data rule | A short scenario course |
| Policy exists, low engagement | Interactive scenarios at the point of use | Reporting rewards |
| High shadow AI use | An approved assistant and gateway visibility | Prompt policies and data controls |
| Regulated data | Role-based training plus data loss controls | Quarterly measurement review |
Conclusion
AI security awareness training works when it changes a few decisions and when controls cover the rest. The UCSD study shows that annual courses do not predict phishing outcomes, while interactive scenarios did better. Pick three behaviors, offer a safe tool, track reporting and shadow AI use, and treat technical controls as part of the program rather than a separate project.
Secure Employee AI Use in Production with NeuralTrust
Give employees an approved route to AI models with policy, logging and data controls built in.
Related Comparisons
- Human-in-the-Loop AI Agents: Oversight That Scales
- ChatGPT Dots: How Agent Security Evolves for Enterprise
- Shadow AI: What It Is, Why It's Dangerous, and How to Stop It
FAQs about AI security awareness training
1. What is AI security awareness training?
AI security awareness training teaches employees to handle AI-related risks: shadow AI tools, AI-enabled phishing and deepfakes, hidden instructions in files and oversharing of sensitive data. It focuses on a few behaviors, gives employees a safe alternative and is measured by what people do rather than by completion rates.
2. Does security awareness training actually work?
Evidence is mixed. A UC San Diego Health experiment with more than 19,500 employees found no significant relationship between recent annual training and phishing failures, and embedded training cut failures by about 2%. Interactive training cut clicking by 19% in relative terms, so format matters.
3. How often should employees take AI security training?
The evidence does not support a fixed interval. NIST SP 800-50 Rev. 1 frames training as a life cycle with ongoing measurement. Use a short onboarding module, add brief scenarios when tools or threats change, and review reporting and shadow AI metrics every quarter.
4. What should AI security awareness training cover?
Cover four situations: using unapproved AI tools, spotting AI-enabled fraud such as deepfake calls, treating untrusted files as possible carriers of instructions, and keeping sensitive data out of external models. Tie each to a clear rule, an approved alternative and a way to report problems.
5. How do you measure awareness training effectiveness?
Measure behavior, not attendance. Track reporting rate, time to report, failure rate per lure type, unapproved AI tool use and blocked sensitive-data events. Avoid copying an industry click-rate target, because failure rates vary widely by lure, from 1.82% to 30.8% in the UCSD study.
6. Can technical controls replace awareness training?
No, but they cover what training cannot. The UCSD authors recommend two-factor authentication and password managers that only work on the correct domain. For AI use, a gateway with prompt policies and data controls limits what a mistaken paste can expose, while training builds the habits that keep those controls working.
About the Author
Roger Howroyd is Head of Global SEO and AI at NeuralTrust, where he leads the company's search strategy across SEO, AEO, GEO, and LLM optimization. He specializes in AI-powered search, content strategy, and SEM. Connect on LinkedIn.
NeuralTrust is the leading platform for securing and scaling AI agents. Named a Pioneer in the Gartner Emerging Market Quadrant for AI Application Security 2026, recognized across six Gartner Hype Cycle reports in 2026, and featured in the Gartner Market Guide for Guardian Agents 2026, the Gartner Market Guide for AI Gateways 2025 and the KuppingerCole Leadership Compass for Generative AI Defense 2025. Headquartered in Barcelona with offices in London and New York. ISO 27001 certified.
Sources
- Ho, Mirian, Luo, Savage, Voelker et al., phishing training experiment at UC San Diego Health, IEEE Symposium on Security and Privacy, 2025: https://sysnet.ucsd.edu/~voelker/pubs/phishtrain-oakland25.pdf
- UC San Diego, "Cybersecurity training programs don't prevent employees from falling for phishing scams," September 17, 2025: https://today.ucsd.edu/story/cybersecurity-training-programs-dont-prevent-employees-from-falling-for-phishing-scams
- NIST, "Building a Cybersecurity and Privacy Learning Program: NIST Publishes SP 800-50 Rev. 1," September 12, 2024: https://www.nist.gov/news-events/news/2024/09/building-cybersecurity-and-privacy-learning-program-nist-publishes-sp-800
- NIST CSRC, "SP 800-50 Rev. 1": https://csrc.nist.gov/pubs/sp/800/50/r1/final
- Verizon, "Vulnerability exploitation top breach entry point, 2026 industry-wide DBIR finds," May 19, 2026: https://www.verizon.com/about/news/breach-industry-wide-dbir-finds
- Malwarebytes, "Americans lost nearly $900 million to AI-powered scams, FBI says," June 8, 2026: https://www.malwarebytes.com/blog/scams/2026/06/americans-lost-nearly-900-million-to-ai-powered-scams-fbi-says
- Help Net Security, "OWASP 2026 LLM Top 10," August 6, 2026: https://www.helpnetsecurity.com/2026/08/06/owasp-2026-llm-top-10-released/
)
)
)
)
)
)
)