Inspiration
We were inspired by the Google Phishing Quiz, which provides users potential phishing emails and they have to determine if they are legitimate.
What it does
We decided to extend this idea and provide an AI-enabled training module where users can put themselves in realistic scam scenarios. We provide a user with 100 coins, which might represent wire transfers of a fiat currency like US dollars or a blockchain transaction of a cryptocurrency, since we anticipate that cryptocurrency transactions may become more prevalent in the future.
How we built it
We used a DeepSeek agent to simulate a scammer attempting to convince a user to transfer some of its 100 coins to its contacts.
Challenges we ran into
We also created our own synthetic dataset using DeepSeek and neural network models to provide a classifier. However, it's very difficult to provide a balanced dataset that is actually representative of real-world input, and our classifier had very little confidence in many actual inputs. We believe this result would have been massively improved by taking advantage of real-world reported messages, which carriers like T-Mobile already receive in large volume every day through SMS spam reporting programs. This kind of dataset would provide enormous detection benefit to check the baseline accuracy of our system and scale it to real-world use.
Accomplishments that we're proud of
Half of our members were first time hackathon attendees. We experimented with new technology and more AI-driven methods of software engineering. We're proud that we were able to simulate a system like this nonetheless.
What we learned
We learned the importance of structured inputs in designing AI-enabled projects, and that their non-deterministic nature makes them very difficult to formally verify. As in any other project, we cannot formally guarantee our agents do not behave in erratic ways, but we can shut things down if they do.
What's next for SignalZero Scam Shield
In the future, we would like to extend it to provide AI-enabled incident response frameworks. We would like to demonstrate the ability to use technology to make more legitimate transactions less urgent and irreversible beforehand, to limit the effectiveness of scam techniques. For instance, with existing technology like e-commerce and credit cards, many ordinary transactions like scheduled grocery deliveries or tax payments could be done through more reversible means, providing a larger window to use this kind of technology to stop fraud attempts during an active incident, rather than focusing mostly on prevention and after-the-fact response.
