The datRail project is building tools that can learn your AI agents' resource, data, and API baselines to automatically enforce runtime guardrails against unintended behavior.
Why datRail?
As AI agents gain autonomous access to local file systems, personal/proprietary data via protocols like MCP, and third-party APIs, traditional static security rules quickly fall short. Agents are non-deterministic by design—predicting every API call, token threshold, or file touch in advance is nearly impossible without crippling their functionality. The datRail project solves this by introducing profile-driven runtime protection: it monitors passively your agent in a control/test run to build a precise baseline of its resource usage and data boundaries, called security profile. Such a security profile is then used to detect deviations during agent runtime. By turning observed behavior into enforceable guardrails, datRail lets developers and power users deploy agentic workflows without giving agents a blank check to their critical data and infrastructure.
Where it runs
Runs on Linux (x86_64 or arm64), on Windows with WSL2, and on a Mac (Intel or Apple Silicon) with Docker Desktop. An agent in an NVIDIA OpenShell sandbox needs native Linux. See platform support (link to the Documents page) for what each one can watch.
Features
A dashboard on your own machine (127.0.0.1) for monitoring your agent's security profile, built from what RailMon sends. Lock a profile as the baseline, and RailDash shows any drift from it, field by field. Monitoring drift from the baseline helps you maintain a guardrail, especially against data leakage: RailDash proposes what the agent is allowed (ports, saved files, uploads, hosts, and tool actions), you review each line, and anything outside it shows as a violation. It also tells you whether the agent can reach any host on the internet or is contained.
Watches an AI agent from the operating system, without changing the agent. Using eBPF, it captures the agent's calls to models and tools, the ports it listens on, and the files it opens, and it scans how the agent is set up every minute. It sends all of it to RailDash as it happens.
Demos
Introduction to datRail
Port drift and guardrail
Tool misuse
Egress Containment
datRail and NVIDIA Openshell