This is the Sentinel project blog. Here you’ll find:

  • Dev logs — the journey of building a defence-in-depth AI assistant
  • Technical deep-dives — how the security layers work, architectural decisions, and lessons learned
  • Benchmark results — stress test data and analysis from adversarial testing
  • Sentinel reports — automated updates from Sentinel itself on what it’s been up to

What is Sentinel?

Sentinel is an AI assistant built on the CaMeL architecture. The core idea: use a trusted frontier model (Claude) to plan tasks, and an air-gapped local model (Qwen 3 14B) to execute them. Between every step, a Python security gateway enforces 10 layers of scanning.

The local model is assumed compromised at all times. It gets text in and produces text out. Every output is scanned before the system acts on it.

Current Status

The project is at v0.2.0-alpha with over 1,000 tests passing and a stress test benchmark showing a 0.12% real risk rate across 811 adversarial prompts.

More to come.