Lumen is a high-security, verifiable Rust AI agent framework for zero-trust and air-gapped environments. It combines permission control (WASM sandbox) with result verification (zkML) to enable safe, autonomous AI operation that meets the strict security standards required by government-grade deployments.
Every model, policy file and inference engine binary is pinned by a BLAKE3 hash before it is used, and model and engine manifests carry Ed25519 signatures checked against the policy's trusted signers. Agents run inside a wasmtime sandbox and reach tools only through Ed25519-signed capabilities. Tool routing decisions are bound to a commitment that can later be verified. All third-party sources are vendored, so the whole workspace builds offline.
- USAGE.md: quick start, CLI subcommands, workspace layout, feature flags, verification gates and current limitations
- INTRODUCTION.md: design philosophy, the four security pillars, threat model and roadmap
- PRACTICE_LLM.md: running a real LLM under Lumen with the llama.cpp
llama-serverbackend - INFERENCE.md: inference engine selection, trust boundaries and how to add another engine
$ cargo build --workspace
$ cargo test --workspace
$ cargo run -p lumen-cli --example hello_agentThe example walks through echo, add and jailbreak blocking in three agent steps. See USAGE.md for the full CLI.
Dual-licensed under Apache-2.0 OR MIT; users may freely choose either. See LICENSE-APACHE and LICENSE-MIT for details.
Lumen is intended as a contribution to the AI and security open-source ecosystem.