Loading...
Loading...
AI agents now act on production financial infrastructure with the same — or greater — privileges as human operators. Most audit systems were built for human speed. This is the trust gap Sentinel was designed to close.
Field Note #1 — Sentinel Series
---
AI is changing fintech faster than most security teams can adapt.
Today, AI agents can review transactions, interact with APIs, assist support teams, and automate critical workflows. But here's the question I keep thinking about:
Can we trust every action they take?
When an AI agent accesses sensitive customer data, approves a workflow, or makes thousands of API requests, financial systems need more than logs — they need accountability.
That's why I'm building Sentinel.
An open-source, event-driven security, audit, and risk intelligence platform for modern financial systems.
The goal is simple: build the trust layer between AI, humans, and financial infrastructure.
---
Financial systems have always had to manage trust. Who accessed what. Who approved which transaction. Who made a change and when.
The tools we built for this — access logs, audit trails, role-based permissions — were designed with humans in mind. A human logs in. A human clicks a button. A human makes a decision.
AI agents don't work that way.
An AI agent can make thousands of decisions per minute. It can access data across systems simultaneously. It can act on behalf of a user in ways that are difficult to trace back to any single intent. And when something goes wrong — when an agent accesses data it shouldn't, or takes an action outside its intended scope — the standard audit trail often can't tell you why it happened, or even clearly that it happened at all.
This is the trust gap. And in financial systems, where the stakes are regulatory penalties, customer data exposure, and transaction integrity, it's not a gap you can afford to leave open.
---
Sentinel is an open-source platform that sits between AI agents, human operators, and financial infrastructure — recording every action, scoring every event for risk, and surfacing anomalies before they become incidents.
It is not a monitoring tool bolted on after the fact. It is designed from the ground up to treat AI agents as first-class actors in a security model — with named identities, behavioral baselines, and the same accountability standards we apply to humans.
The core capabilities:
Immutable audit ledger. Every action — human or AI — is recorded with a cryptographic signature. Records cannot be modified after the fact. A compliance auditor can verify, months later, that the log they're reading is exactly what was written at the time.
AI actor identity. AI agents are named entities, not anonymous service accounts. Every credential issued to an AI agent carries the agent's name and version. Every action that credential takes is attributed precisely to that agent.
Behavioral baseline scoring. Risk signals compare an actor's current behavior to their own historical pattern — not a global threshold. A reconciliation bot that normally processes 10,000 records per hour is judged differently from a support bot that normally handles 20 API calls per minute. A spike is only meaningful relative to what's normal for that actor.
Real-time alerting. Risk scores are computed within seconds of each event. Alerts fire immediately when a score crosses a configured threshold. A security team can respond to an anomaly while it's still happening, not three days later during a log review.
Compliance reporting. Audit evidence packages show AI actor attribution as the headline — which named AI agents were active, what they accessed, how their behavior compared to their baseline. The kind of breakdown a regulator needs to evaluate whether your AI systems operated within appropriate scope.
---
Security infrastructure for AI systems is an unsolved problem. Most organizations are either ignoring it or bolting on tools designed for human-centric threat models and hoping it's close enough.
It isn't.
The threat model for AI agents is fundamentally different. The signals that matter are different. The attribution requirements are different. The compliance questions are different. And the speed at which an AI agent can cause damage — compared to a human making manual decisions — means the detection-to-response gap has to be measured in seconds, not hours.
Building Sentinel in public is a way to work through these problems in the open, with feedback from engineers who are facing the same challenges. Every architectural decision gets documented. Every tradeoff gets explained. Every phase builds on the last.
Over the coming weeks, I'll be documenting every phase: from event-driven design and immutable audit logs to observability, risk scoring, and the dashboard a security team actually uses during an incident.
If you're working on Python, Django, distributed systems, or fintech infrastructure — I'd love your feedback as the project evolves.
---
Sentinel is open source. GitHub: [github.com/Gwerdonatus/Sentinel](https://github.com/Gwerdonatus/Sentinel)
Next: Building Sentinel Phase 1 — Laying the Foundation