ProofPath

Inspiration AI has made impersonation scams much harder to recognize. Scammers can now copy someone’s writing style, clone voices, generate realistic images, and create convincing messages very quickly. Most scam-detection tools try to answer one question: “Does this look like a scam?” We wanted to approach the problem differently. Even if a suspicious message looks completely real, the more important question is: what is this person asking me to trust, and can I verify it before I act? That idea became ProofPath. Our goal was to create a system that does not simply give users a scam probability, but instead breaks suspicious interactions down into the claims, actions, and trust assumptions behind them.

What it does ProofPath analyzes suspicious messages and helps users understand exactly what needs to be verified before they take action. For example, if someone sends: “Hey, my phone broke. This is my new number. Can you send me €500 urgently?” ProofPath identifies things such as: who the sender claims to be what action they want the user to take changes or inconsistencies, such as a new phone number or payment destination risk signals such as urgency or financial requests which important claims are still unverified Instead of simply saying “this is probably a scam,” ProofPath shows the user what is still unknown and what should be verified before they act. Our main idea is simple: verify before you trust.

How we built it We built ProofPath as a web application so it could be simple to use on both desktop and mobile devices. The system follows a structured process. The user enters a suspicious interaction, and the AI analyzes the message to extract the important information. It identifies the claimed identity, requested action, contextual warning signs, and any trust gaps that may require verification. The results are then shown in a clear interface that separates what is known, what looks unusual, and what still needs to be verified. One of our main goals was to avoid treating AI as if it always knows the truth. We use AI for what it is good at, understanding messy human communication, while keeping verification separate from prediction. We also focused heavily on the user experience. Cybersecurity tools can quickly become confusing, so we designed the product around a simple flow: Understand → Identify trust gaps → Verify

Challenges we faced One of our biggest challenges was deciding how much responsibility the AI should have. At first, we considered building a more traditional scam detector that would analyze a message and return a suspiciousness score. However, we realized that a result like “91% likely to be a scam” can create a false sense of certainty. A sophisticated scam may look completely legitimate, while an unusual but real message may look suspicious. Because of this, we shifted the project away from prediction and toward verification. Another challenge was turning AI output into something genuinely useful. Raw AI responses can easily become long and confusing, so we had to structure the output into information that a user could understand immediately. We also had to balance our ideas with the limited time available during the hackathon. There were many features we wanted to explore, including deeper behavioral analysis and more advanced verification methods, but we chose to focus on making the core idea work properly instead of building too many unfinished features. The interface was also a challenge. Earlier versions were too cluttered and contained too much information at once, so we repeatedly simplified the design until the main question was always clear: what still needs to be verified?

What we learned The biggest thing we learned is that cybersecurity is not always about building a better detector. AI-generated content will continue becoming more realistic, which makes it increasingly difficult to determine whether something is real simply by analyzing how it looks or sounds. Instead of trying to perfectly separate real from fake, security systems can reduce risk by requiring stronger evidence before important actions are taken. We also learned how important explainability is when using AI for cybersecurity. A user should understand why something has been flagged and what they should do next instead of only receiving an unexplained score. From a technical perspective, we gained experience using AI to turn unstructured language into structured information, designing risk and verification workflows, connecting frontend and backend systems, building a responsive web application, and collaborating through Git and GitHub. Most importantly, the project changed the way we thought about the original problem. Instead of asking, “Can AI detect the scam?” We started asking, “Can we make the scam unsuccessful even when the AI cannot tell that it is fake?” That became the foundation of ProofPath.