A cybersecurity question answering assistant that motivates users to follow its advice

It is estimated that more than 90 percent of cybersecurity incidents can be attributed at least in part to human error, typically someone failing to make the right decision. These errors in decision making can range from individuals clicking a suspicious link in an email, to failing to use a VPN, to not taking the time to install anit-malware on their respective devices.

For cybersecurity professionals, training people to make better decisions has also become increasingly challenging, as there is a surfeit of information that people would need to have to help them make the right decisions. When in doubt, people have traditionally relied on search engines and social media sites like Reddit for advice, and increasingly they are now also turning to chatbots. But how good is the advice provided by these chabots?

A team of Carnegie Mellon University researchers led by Norman Sadeh, professor in the Software and Societal Systems Department, has been looking at this issue, collecting a large number of common cybersecurity questions submitted by people in the context of their daily lives and studying the answers returned by state-of-the art chatbots such as ChatGPT4.

“We were surprised to find that the answers returned by state-of-the-art chatbots are typically quite accurate,” said Sadeh. “At the same time, our study found that these answers are often not terribly effective. They are often hard to understand, short on concrete actions users can take and, above all, they fail to motivate people to follow recommendations.”

The team proceeded to study prompt engineering techniques that could coax large language models (LLMs) to generate more effective answers to everyday cybersecurity questions. In their work, they drew on Protection Motivation Theory and Self-Efficacy Theory to construct prompts that lead LLMs to produce significantly more effective answers.

Protection Motivation Theory emphasizes the need to highlight risks as a way of motivating people. Accordingly, prompts used by Sadeh and his team were designed to elicit answers that emphasize the risk of not following recommendations.

The result of this research was implemented in the form of a Security Question Answering (QA) Assistant deployed as a Google Chrome extension. In their most recent study, which they presented earlier this week at the Usable Security and Privacy Symposium (USEC) 2025 in San Diego, Sadeh and his team reported on a pilot involving 51 people who used their Security Assistant as part of their regular, daily activities over a period of 10 days. On average, these people asked more than two questions per day, resulting in the collection of more than 1,000 questions covering a diverse number of topics. Each evening, participants were asked to review the answers they had received and help evaluate their effectiveness. Participants were divided into two groups: one with the benefit of the prompt engineering technique and one without. 

“We were surprised to see how effective our technique was with real people in the context of their regular everyday activities,” said Lea Duesterwald, an undergrad researcher working on Sadeh’s team and the lead author on the paper.

Security is known to be a secondary task, with people ignoring recommendations as they continue to focus on their primary tasks.

Norman Sadeh, Professor, CMU Software and Societal Systems Department

“The results showed increases in effectiveness along all dimensions, with most of these increases being statistically significant,” said Ian Yang, a research assistant in CMU’s Software and Societal Systems Department who also contributed to the study. “In particular, our technique was shown to significantly increase the likelihood that people follow recommendations contained in the answers generated by the Security Assistant.”

The resulting paper, “Can a Cybersecurity Question Answering Assistant Help Change User Behavior? An In Situ Study,” details the findings of what the authors believe to be the first in situ evaluation of a cybersecurity QA assistant. 

The study found that participants who received prompts were more likely to understand and act upon the advice, and they found the answers more helpful. The vast majority of participants indicated that, if given access to such a cybersecurity assistant, they would use it with as many as a third of participants indicating they would likely use it everyday or at least several times per week.

“These findings are exciting, because they indicate that we can really change people’s behavior,  get them to follow recommendations  and improve their security posture,” said Sadeh. “Security is known to be a secondary task, with people ignoring recommendations as they continue to focus on their primary tasks. Getting people to actually follow recommendations in cybersecurity is a really big deal.”

The team is continuing its research and looking at additional opportunities to further improve the effectiveness of its assistants, including exploring opportunities to personalize answers.

“Ultimately, we would like to go beyond just answering people’s cybersecurity questions and enable our assistants to pro-actively intervene rather than wait for users to ask questions,” said Sadeh.