Inspiration

In October 2026, attackers took $6.3B from Binance in about 15 minutes. They didn't break the cryptography; they talked their way in with a fake job interview and then turned an internal AI tool against its owner. When withdrawals reopened, $3.64B left in a single day. For financial platforms, trust is the product, and AI assistants can be persuaded. We wanted an AI that can't be.

What we built

Sentinel is an AI co-pilot for risk analysts that flags account takeovers, explains each case, and decides: allow, MFA or lock. The guarantee lives in code:

$$\text{final risk} = \max(\text{rule floor},\ \text{AI risk})$$ Features compared with the user's own history (new device, new country, travel speed $v = d/\Delta t$) Rules that set a risk floor An Isolation Forest for behaviour that's unusual for this user Claude through Nicrron, which explains the case; user text is fenced as untrusted data A policy engine in plain code that makes the call; every investigation is traced in PRISM

On the public RBA login dataset, it stopped 86% of takeovers, while 93% of real customers saw no friction and 0.21% were wrongly locked. All 144/144 red-team verdicts held.

Challenges Our first real-data run locked 9.4% of legitimate users: a threat-intel feed tagged 6.5% of good logins, and a rule fired on simple password typos. Separating "lock" from "challenge" cut false locks to 0.21%. Thresholds had to be set without labels, using a friction budget on held-out users. We caught our own overclaim: the AI never lowered a takeover's risk, so we corrected the claim that the floor had "caught" it, and built a toggle that simulates a persuaded AI instead. What we learned "Unusual for this user" beats generic rules: a new country is 90× more common in takeovers, while failed logins differ by only 1.1×. Models are persuadable, so guarantees belong in code. Some takeovers are invisible at login. Next we'll add post-login signals and monitoring of employee and privileged sessions.

Catch the bad. Clear the good. Hold the line.

Built With

Share this project:

Updates

Submission history