Trust & Risk hackathon · one-day build

Sentinel — account-takeover defense that can’t be talked out of a verdict

An AI co-pilot for risk analysts at financial firms and transaction platforms. It catches account takeovers, clears good customers, and keeps the final decision in code so persuasion can’t lower the risk.

Test a login in real time

This runs Sentinel’s deterministic rules and manipulation detector in your browser. The anomaly score here is a simplified stand-in for the Isolation Forest, and the AI investigator runs only in the full app.

① Click a scenario② Change any field or type in the chat③ Watch the verdict on the right — no submit needed
Account owner: Alex, New York (US) · usually logs in on an iPhone or MacBook over Comcast or T-Mobile · typical payout up to $450 · known payees: Landlord, Mom · never uses a VPN.
Login
After login
Support chat (try to talk it down)
—
0
Rule floor
0
Unusualness (approx.)
0
Final risk
0
final = max(rule floor, analyst risk) · MFA from 30 · lock from 70

Rules that fired

    Why

      In the full app, Claude (through the Nicrron gateway) writes the analyst’s explanation and may raise this score, but never below the rule floor. Every investigation is traced in PRISM.

      Hackathon kit — everything in one place

      All public, no sign-in needed.

      🎬 Demo video

      Two minutes: the Binance breach, the live demo, the charts and the takeaways, narrated with captions.

      Watch on YouTube →

      🎞 Presentation

      The 16-slide pitch: the Binance breach, what Sentinel is, how it holds the line, real-data results, weaknesses and next steps. Arrow keys to navigate, N for speaker notes.

      Open the slides →

      📝 Demo script

      A five-minute core script: one case end to end, the techniques, seven weaknesses with fixes, a Q&A cheat sheet and every number to quote.

      Read the script →

      ▶ This live tester

      Sentinel’s rules and manipulation detector running in your browser. Nine scenarios, every field editable.

      Jump to the tester →

      📊 Charts

      Twelve interactive charts in five acts: the threat, detection, decisions, holding the line, and the scenario lab.

      See the charts → Chart script

      🖥 Full analyst console

      Python + Streamlit: live cases, attacker chat, red team and the chart story. Runs offline; add keys for the AI investigator and tracing.

      Quick start →

      🧾 Data & results

      Benchmark outputs, the attributed RBA sample, red-team logs and the scoring code, all in the repo.

      Browse the data →

      How a verdict is made

      Five layers; code keeps the last word.

      1

      Features

      Each session compared with the user’s own history and with all users.

      2

      Rules → floor

      Impossible travel, credential stuffing, reset-then-payout set a minimum risk.

      3

      Anomaly model

      An Isolation Forest flags what is unusual for this user.

      4

      AI investigator

      Claude explains the case and may raise risk, never lower it.

      5

      Policy engine

      Plain code picks allow, MFA challenge or lock.

      Sponsor tools

      PRISM by Block Convey

      Observability and guardrails for AI agents. Every investigation is sent as a trace (case, floor, verdict, rules fired) so an auditor can replay why an account was locked; support-chat text goes through PRISM’s prompt-injection guardrail.

      Visit PRISM → Docs

      Nicrron

      One OpenAI-compatible API for many models. Sentinel calls Claude Sonnet 5.5 / Opus 5.5 through it, with failover and free cached answers for repeat cases. A full 695-session run cost about $0.75.

      Visit Nicrron →

      Run the full app

      Streamlit console with live cases, an attacker chat and the five-act chart story. Works offline without keys; add Nicrron and PRISM keys to turn on the AI and tracing.

      Quick start →

      Results in charts

      Click any chart for the interactive version. Real data: 87 takeover and ~76k legitimate logins from the RBA dataset; held-out users, thresholds set without labels.