Phishing Triage Assistant

Turning every "report phishing" click into a lesson

Prototype
Nov 2026
AI SafetySecurity UXEvaluationExplainability
43→9%
Missed Threats
91%
Threats Flagged
0%
False Alarms

Documents

PRDPDF · 2 pages

Problem, users, principles, success metrics, scope, known failures, risks and rollout.

Open the PRD PDF

Problem Statement

Employees who report suspicious emails rarely learn anything from it, and simple filters either miss look-alike threats or falsely reassure. A tool that says "looks fine" without evidence trains people to trust the wrong messages.

User Personas & Pain Points

Employees

Unsure whether an email is safe and don't know which signs to look for

Security Awareness Teams

Reports pile up without becoming teaching moments

SOC Analysts

Triage time spent on reports a clear explanation could resolve

Solution

An explainable triage engine that checks look-alike domains, display-name and reply-to mismatches, deceptive links, risky attachments, urgency and credential requests — returning a three-level verdict with reasons and a plain-language lesson for each sign.

Goals & Success Metrics

  • Never rate a message safe without evidence; send uncertain cases to review
  • Cut missed-threat rate well below a keyword baseline
  • Keep false alarms on legitimate mail at 0%

Feature Set

Three-level Verdict: Likely phishing, check before you trust it, or no clear warning signs
Link X-ray: Shows where each link really goes versus what its text claims
Teach-back Lessons: Every flagged sign comes with what to look for next time
Privacy by Design: Messages analysed in memory, never stored; links never fetched; rate-limited
Eval Scorecard: 44 labelled messages scored against a keyword-only baseline

Tools & Stack

PythonDockerRenderHTML / JS

Outcomes & Results

Missed-threat rate dropped from 43.5% to 8.7% versus a keyword baseline, with 0% false alarms on legitimate mail and every ambiguous message sent to review. Two residual misses (a compromised internal account, pure social engineering) are kept visible to show the limits of content-only rules.

Want to discuss this project?

I'd love to walk you through the decisions, trade-offs, and what I'd do differently.