About
An open-source project by Keter Labs
Reality Kernel puts the decision about what an AI agent may do where the agent cannot argue with it: in the Linux kernel.
The project
Agents now run shell commands, read files and open network connections with real privileges. Guardrails that inspect prompts or tool calls help, but a prompt injection, a bug or a child process can step around them. Reality Kernel enforces policy at the system call instead, with eBPF programs attached to the kernel's BPF-LSM and cgroup hooks, and records every decision in a signed audit log.
The project is developed in the open under the MIT and Apache-2.0 licences, with the BPF programs dual-licensed GPL-2.0-only or MIT. Every capability it claims is tied to named tests that run on real kernels in CI, and the evidence page lists them.
History
Reality Kernel began as a hosted service that checked agent commands through an API. In 2026 it was rebuilt from the ground up as kernel enforcement and released as open source. The hosted API and its SDK have been retired.
Keter Labs
Keter Labs, founded in 2026, is a deliberately small research outfit working on adversarial and complex systems. Reality Kernel is its first open-source project.
The industry is racing to build autonomous systems, but trying to secure them with language models is mathematically flawed. Agent boundaries must be deterministic, built in Rust, and enforced offline.
Contact
General questions: contact@realitykernel.dev. Security reports: security@realitykernel.dev, or see the security page. Bugs and feature requests: GitHub issues.