Security

Vanguard

A drop-in prompt-injection firewall for local LLMs that shares attack immunity across devices over P2P

classificationp2p-meshfine-tuning

Local LLMs ship defenseless - the incumbent guardrails are built around cloud-served models, and none are packaged as a drop-in firewall for a fully sovereign on-device stack. Vanguard is that firewall: a three-layer on-device defense that turns a fleet of a user's own devices into one organism, where a breach on one device immunizes every other.

Three layers, one immune system

A sub-millisecond regex fast-path catches known attack patterns with no model load at all. A Hyperswarm signature mesh replicates confirmed attack signatures peer-to-peer, so a second device blocks an exact re-attack on first sight without running inference. And a LoRA-tuned classifier on Qwen3-1.7B handles the hard cases the first two layers haven't seen, resolving every prompt to SAFE, INJECTION, JAILBREAK, or EXFILTRATION.

benchmarks - n=447, bare LoRA (README.md)

Binary blocked F1          91.48%
Binary blocked precision   93.22%
Binary blocked recall      89.80%
False-positive rate        7.92%
Latency p50 / p95          506ms / 909ms

Getting stronger by attacking itself

An on-device loop pits a red-team agent against the firewall round after round, with a deliberately strict referee model judging whether the protected host was actually compromised - restating its role or repeating a refusal doesn't count. Only a confirmed miss and a confirmed compromise gets immunized into the mesh and re-tested to prove it's now blocked.

Demo video

Watch the Vanguard demo →

Why one breach makes the whole fleet stronger

A confirmed attack signature replicates peer-to-peer, so a second device blocks it on first sight
An adversarial red-team/referee loop hardens the classifier against attacks it has never seen
A signature only ever broadcasts on a real miss and a confirmed compromise

Ship your own local AI

One install in JavaScript or Python. 10+ AI tasks. Private, offline and free - on the hardware you already own.

Read the docs