Runtime Behavioral Security for AI Agents

The pause between intent and action.

Sentinel intercepts every agent action at two deterministic checkpoints, running four independent defense layers before anything executes. Ships as an npm package. Works with Claude Code today. No LLM in the monitoring pipeline.

Get started with Claude Code
npm install @tuent/sentinel npx sentinel init claude-code
Two commands. Every tool call passes through policy enforcement.
Architecture Advisor
"If I had spare time, I would build that."
Advisor · Sr. Director of PM, AI Security · Palo Alto Networks
Additional Ongoing Executive Advisory From
Google Cloud
80%+
Attack success rate for behavioral control traps across five production agent frameworks.
Shapira et al. (2025), cited in Franklin et al., "AI Agent Traps," Google DeepMind, 2026.

Four layers. No model in the loop.

Every action passes through a deterministic gate. Same input, same verdict. Replayable. Auditable.

Agent Action 01 Role Boundary Policy + 26 sensitivity rules Credentials · System · Config 02 Intent Action matches task? Zero LLM cost 03 Baseline Learned behavior 7 deviation checks 04 Hooks Customer rules Block before exec Allow Block DETERMINISTIC · SAME INPUT, SAME VERDICT · NO LLM IN THE PIPELINE
01

Role Boundary

Action allowed by policy? 26 built-in sensitivity rules across credentials, system files, databases, and config.

02

Intent Alignment

Does the action match the stated task? Zero LLM cost. Five consecutive misaligned actions escalate to critical.

03

Behavioral Baseline

Seven deviation checks against learned patterns. Catches anomalies invisible to static rules.

04

Hook Prevention

Customer-defined enforcement. When a hook returns Block, the action never executes.

Built, not pitched.

2,833
Tests
~85,400
Lines TypeScript
5+
Frameworks
0.42%
FP Rate
$0
Runtime Cost

Live on npm. Production Claude Code integration shipped. Two commands to enforcement. Listed on the AARM Builders Registry. Enterprise validation: 238 actions, zero false blocks. Ed25519-signed audit trail for SOC 2, GDPR, HIPAA, EU AI Act.

Design Partner
Omniloy (Health AI)
Standards Body
AARM Builders Registry
James Cunningham
Co-founder, CEO

GTM, partnerships, fundraising. Founding CTO of a prior angel-backed AI venture.

Charlie Greenberg
Co-founder, CTO

Primary architect. Built the enforcement engine and every line of the test suite.

Full architecture in the technical brief.

Defense layers, compliance, enterprise validation, live demo walkthrough.

See it catch a real attack.

Live demo. Real Claude Sonnet 4 agent. Real prompt injection. All four layers catch it independently.

Request Demo

Or get started now: npm install @tuent/sentinel

Technical Brief

Full architecture documentation.

Defense layers, compliance, enterprise validation, live demo walkthrough.

Request Demo

See Sentinel catch a real attack.

Live demo with a real Claude Sonnet 4 agent. We run it on your screen.