Safely

Inspiration

Safely was inspired by a problem that feels personal for a lot of people: phishing is not just a cybersecurity issue, it is a family issue. Many of us know someone, especially an older parent or grandparent, who has received a fake bank alert, delivery message, or urgent email that looked real enough to trust. We wanted to build something that could step in at that exact moment of uncertainty and help people make a safer decision before they clicked, logged in, or gave away sensitive information.

We were especially motivated by the idea that existing warnings are often too technical, too late, or too easy to ignore. We wanted Safely to feel more human, more understandable, and more helpful for everyday users who are not security experts.

What it does

Safely is a Chrome extension that analyzes webpage content in real time and warns users when a page shows signs of phishing, impersonation, or scam behavior.

It looks for signals such as:

suspicious or misleading domains
fake login or sign-up pages
brand impersonation
pressure tactics and urgent calls to action
risky webpage structure and suspicious content patterns

Once Safely detects enough suspicious behavior, it presents a clear warning to the user and explains why the site may be dangerous in plain language. We also added AI-generated voice narration so the extension can speak warnings aloud, making the product more accessible and more engaging during the user experience.

How we built it

We built Safely as a Chrome extension using a modular TypeScript-based architecture. The main components included:

a content script to scan webpage content and extract suspicious signals
a background service worker to handle scan flow, scoring, and coordination
a popup UI to display verdicts, explanations, and controls
a risk scoring engine to combine suspicious features into a final assessment
Google Safe Browsing as one external reputation signal
generative AI APIs for clearer explanations and voice narration

Our process was roughly:

detect and extract webpage features
score those features for phishing and scam risk
enrich the result with external checks like Safe Browsing
show the user a readable verdict and explanation
optionally narrate the warning out loud

A big part of our work was making the system more explainable. We did not want the extension to just say a page was “bad.” We wanted it to tell the user things like: the domain looks suspicious, the page is pretending to be a known brand, or the language is pressuring the user to act quickly.

Challenges we ran into

One of the biggest challenges was balancing sensitivity and accuracy. It was easy to make the extension flag too many things, but much harder to make it flag the right things for the right reasons.

For example, we ran into cases where:

brand detection was too broad and triggered just because a known company name appeared somewhere in the text
urgent language detection was too naive and flagged phrases like “security breach” even when they were not actually pressuring the user to do anything
some phishing domains were not being caught by external reputation checks even though they were clearly suspicious
AI voice narration had edge cases such as duplicate playback, delayed playback, and stop-button behavior that needed tighter control
microphone permissions and speech interaction needed smoother state handling so the user did not have to keep re-approving access

Another challenge was keeping the product polished under hackathon time pressure. It is one thing to make a detection pipeline work, but another to make the extension feel like a real product with a clean UI, understandable warnings, and a demo-ready experience.

Accomplishments that we're proud of

We are proud that Safely became more than just a technical scanner. It became a product with a clear user story and real-world value.

Some accomplishments we are especially proud of:

building a working Chrome extension instead of just a mockup
creating a real-time webpage scanning pipeline
combining heuristic detection with external reputation checks
adding AI-generated voice narration for accessibility and impact
designing explainable warnings rather than black-box outputs
focusing on a human-centered problem that judges and users can immediately understand
shaping the product around protecting everyday people, not just power users

We are also proud of how much the project evolved. What started as a scam detection idea became a more thoughtful safety tool centered on clarity, trust, and prevention.

What we learned

We learned that phishing detection is as much a product-design problem as it is a technical problem. A system can be technically impressive and still fail if users do not understand the warning or do not trust it.

We also learned:

simple keyword matching is not enough for reliable phishing detection
reputation feeds are useful, but they are incomplete on their own
explainability matters a lot for user trust
real-time browser tools need careful state management to avoid duplicate actions and race conditions
accessibility features like voice can make a security product feel much more usable and memorable
a good hackathon project needs both technical depth and emotional clarity

Most importantly, we learned that cybersecurity tools are stronger when they are built for real human behavior, not just technical correctness.

What's next for Safely

Next, we want to make Safely more accurate, more personalized, and more widely useful.

Our roadmap includes:

improving suspicious-domain and impersonation detection
expanding beyond webpages into email and SMS scam analysis
strengthening threat intelligence and domain reputation checks
reducing false positives with more contextual scoring
improving conversational voice interaction
adding an accessibility mode designed for older adults and less technical users
building family-oriented safety features so people can better protect loved ones online

Long term, we want Safely to become a trusted digital safety companion that helps people pause, understand risk, and stay protected before one wrong click becomes a costly mistake.