Inspiration

Last year, Americans lost over $16 Billion to cyber fraud and internet crimes. (FBI). Despite this figure being up 33% y/y, we expect it will only continue to grow, with sophistication improvements fueled by AI.

What it does

Ward is a cutting-edge security extension that leverages the Chrome Prompt API to proactively scan the current page for threats and remediate with actionable next steps.

How we built it

  • Claude Code/Sonnet 4.5
  • Chrome Prompt API
  • Live phishing content (personally received by cofounder) for training data
  • https://abuse.ch/ for additional training data
  • Gemini Nano for tool logo

Challenges we ran into

  • Prompting: underlying prompts were iteratively updated as we expanded Ward's scope to include gray area threats and email inbox content
  • Accommodation of edge cases regarding content extraction, such as iframes
  • Scalability: Midway through the project, we noticed that Ward initially failed to run in multiple tabs simultaneously; we fixed this by adding concurrency
  • UI/UX were high priority, as we initially selected a subset of the population age-wise for our ideal customer persona. We sought to avoid overloading or overwhelming the end user (for example, we considered inserting an upward facing arrow beneath the plugin icon in the browser toolbar, to guide users to a potentially unpinned copy of Ward in order to view our long-form findings, but we ultimately decided against it, as we felt it risked delegitimizing the quality of the banner, potentially leading to drop off)

Accomplishments that we're proud of

  • UI/UX maximizing inevitable constraints
  • Gray area detection of adware, fringe threats, investment scams

What we learned

  • The art and science of popup mechanisms: A great deal of effort went into design, as we considered the end users' POV for alert banner, sizing, and content (with the goal of avoiding user drop off)
  • We experimented with multimodal inputs at first, as we initially thought they may be promising, but later found that text was the most viable for our immediate needs, and plan to pursue multimodal inputs later on

What's next for Ward

  • Multimodal Inputs
  • Blocking
  • False Positive Reporting/Intake Mechanism
  • Support for an additional 12 languages
  • "Easy" Mode for youth and the elderly

Built With

Share this project:

Updates