# PrismGuard

One canonical identity. Human page: https://www.insightits.com/products/prismguard.html

| Field | Value |
|-------|-------|
| Slug | `prismguard` |
| Status | published |
| Family | aiSecurity |
| Infra category | AI Security |
| Record type | product |
| Parent | — |
| JSON | https://www.insightits.com/catalog/prismguard.json |
| Markdown | https://www.insightits.com/catalog/prismguard.md |
| GitHub | https://github.com/insightitsGit/PrismGuard |
| PyPI | https://pypi.org/project/prismguard/0.1.10/ |
| Install | `pip install "prismguard[prism,guard-model]==0.1.10". Apache-2.0 open core — free forever. Optional ops: ChorusControl Enterprise (CONTROL).` |
| Version | 0.1.10 |

## What it is

Self-hosted prompt-injection firewall and ChorusGraph security plug-in. Returns auditable allow/block with resolution_gate and decision_source. Opt-in profiles: web_chat, light/heavy ONNX, and 0.1.10 domain_pilot after train-first.

## Problem

Opaque guards and scanner-only stacks make it hard to audit which layer blocked a prompt, and cold holdout / domain traffic still needs a local, fail-closed decision.

## Alternatives

- **LLM Guard (PromptInjection scanner)** — Published law-domain holdout compares a configured PrismGuard pipeline to one LLM Guard PromptInjection scanner configuration. Feature-asymmetric; not a whole-product guarantee. ([source](https://www.insightits.com/compare/prismguard-vs-llm-guard.html))

## Features

- **Auditable Allow/Block** — Every decision exposes resolution_gate and decision_source for compliance logs — not a black-box score.
- **Self-Hosted Firewall** — Sits in front of the LLM (and can scan assistant output). Local ONNX classifier; no OpenAI required by default.
- **ChorusGraph Security Plug-in** — make_guard_handler before RAG/LLM. Domain-agnostic for legal, healthcare, and finance — align ONNX with shadow + feedback on your traffic (law pack is published proof).

## Architecture

Layered local rules + taxonomy + optional ONNX + optional judge escalation. Every decision reports resolution_gate and decision_source. ChorusGraph wiring: make_guard_handler + route_after_guard before cache/RAG/LLM.

## Use cases

- Standalone prompt screening before RAG or LLM
- ChorusGraph security port for legal, healthcare, and finance traffic
- Train-first domain_pilot after you have labeled traffic

## Benchmarks

In the published 14-attack law holdout, the configured PrismGuard pipeline blocked 14/14 while the measured LLM Guard scanner blocked 9/14; both allowed 25/25 normal holdout cases. Separately, finance mid bake-off shows Prism pack PI 100% vs peer 85% vs AgentCore 45% (task ties AgentCore at 100%).

### Disclosures

- Domain-aligned integrated firewall pipeline versus one LLM Guard prompt-injection scanner configuration (law). Separate finance mid bake-off versus LangGraph+LLM Guard and AgentCore Runtime.
- PrismGuard uses its law overlay, authored seed, ONNX artifact, and selective escalation on the law holdout. Finance mid uses domain-calibrated finance artifact + pack wiring. Results are not model-only or whole-product guarantees.
- Broadly beats LLM Guard on detection
- Universal prompt-injection accuracy
- Beat AgentCore / AWS overall
- Cross-host latency win vs AgentCore
- Mean LLM efficiency win vs AgentCore
- Enterprise healthcare or finance certification
- Performance superiority over Lakera Guard or NVIDIA NeMo Guardrails
- Enterprise-ready without a design-partner pilot

### Evidence URLs

- https://github.com/insightitsGit/PrismGuard/blob/master/benchmark/law/results/current/COMPARISON_REPORT.md
- https://github.com/insightitsGit/PrismGuard/blob/master/benchmark/law/results/current/COMPARISON_REPORT.md


## GitHub

https://github.com/insightitsGit/PrismGuard

## PyPI

https://pypi.org/project/prismguard/0.1.10/

## Documentation

- [Prompt-injection firewall guide](https://www.insightits.com/guides/prompt-injection-firewall.html)
- [PrismGuard vs LLM Guard](https://www.insightits.com/compare/prismguard-vs-llm-guard.html)
- [LLM security guardrails guide](https://www.insightits.com/compare/llm-security-guardrails.html)
- [Design notes](https://github.com/insightitsGit/PrismGuard/blob/master/docs/prismguard-design.md)
- [Integration guide](https://github.com/insightitsGit/PrismGuard/blob/master/docs/integration-guide.md)
- [https://github.com/insightitsGit/PrismGuard/blob/master/benchmark/law/results/current/COMPARISON_REPORT.md](https://github.com/insightitsGit/PrismGuard/blob/master/benchmark/law/results/current/COMPARISON_REPORT.md)
- [enterprise](https://github.com/insightitsGit/PrismGuard/blob/master/docs/enterprise-product-model.md)
- [userUpdates](https://github.com/insightitsGit/PrismGuard/blob/master/docs/user-updates.md)
- [ci](https://github.com/insightitsGit/PrismGuard/actions/workflows/ci.yml)
- [onnxRelease](https://github.com/insightitsGit/PrismGuard/releases/tag/v0.1.2)
- [chorusgraphGithub](https://github.com/insightitsGit/ChorusGraph)

## Is not

- Enterprise certification or SOC2
- A broad “beats LLM Guard on detection” claim
- A performance claim against Check Point AI Guardrails or NVIDIA NeMo
- Healthcare/finance domain packs with accuracy guarantees (use domain_pilot + train on your traffic)

## Related products

- [chorusgraph](./chorusgraph.md)
- [prismshine](./prismshine.md)
- [prism-pack](./prism-pack.md)