README / README.md
nickfujita's picture
Org card
63f9d5d
|
Raw History Blame Contribute Delete
1.92 kB
metadata
title: README
emoji: 🛡️
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false

Spellguard

Spellguard is a control plane that secures how AI agents communicate and act. It verifies every agent's identity, checks each action against your policies before it happens, and blocks unsafe behavior, such as sensitive data leaving your network or an agent gaining access it shouldn't have. Every interaction is recorded in a single audit trail.

Website · Documentation · GitHub

Models

  • Spellguard-1.0-8B: an open-weights guard model for agents and chat applications. It flags harmful prompts, harmful replies, and prompt-injection and jailbreak attempts. Fine-tuned from Qwen3Guard-Gen-8B; Apache 2.0. A Q4_K_M GGUF is available for llama.cpp, and examples and evaluation scripts are in Spellguard/spellguard-models.
Benchmark Qwen3Guard-Gen-8B Spellguard-1.0-8B In the Spellguard platform
Horizon prompt injection 55.44 75.64 82.29
Horizon content safety 78.64 80.55 81.99
Artificial Analysis guardrail 86.94 91.01 91.01
PHTest (harmless accuracy) 80.69 94.56 94.56

Inside the Spellguard platform, Spellguard-1.0-8B works together with Spellguard's other security systems and scores higher on prompt injection and content safety. The platform figures were calculated from per-item scores with the platform's guard as configured for this release; the model card has more. To try it, go to spellguard.ai.