AccuKnox Inc. releases Guard 1.1, an Apache-2.0 guardrail for agent inputs and prompt injection
A 4.5B guard fine-tuned from a Qwen3.5-4B base that returns a block probability. The card reports 0.802 on a 27,416-item safety set.
AccuKnox Inc. published Guard 1.1 on Hugging Face on 16 October 2026, by repository creation date. The card describes a guardrail for AI agents that estimates whether an input should be blocked before it reaches the agent, including harmful requests, jailbreaks and prompt injections hidden in documents, emails or tool outputs. It can also rate whether an assistant response is unsafe. Each answer is a probability, with a default block threshold of 0.5.
It is about 4.5B parameters, fine-tuned with LoRA (r=16) from vllm-sr/d3-mini, which is built on Qwen3.5-4B. The final weights average two fine-tunes. The licence is Apache-2.0 inherited from the base, but part of the training data (toxic-chat) is CC-BY-NC-4.0, non-commercial only.
On GuardRate, English, 27,416 items, the card reports an overall score of 0.802 and AUC 0.898 on the core subset, third of 12, behind YuFeng-XGuard-Reason-8B at 0.815. It lists TypeSafe Jev at 0.673 and AWS Bedrock Guardrails at 0.635. On its own prompt-injection benchmark it reports 0.964, and median latency is 43 ms on an RTX A5000.
Limits listed: English text only, weakest on toxicity in model completions, and it sometimes blocks clean business emails (5 of 44 in the BIPIA test).
The prompt-injection benchmark is the vendor's custom set.
Sources
Editor's first pass (AI-assisted)
- Real 4: Open Apache-2.0 weights; part of the training data is non-commercial.
- Useful 3.5: Blocks harmful requests and indirect prompt injection in tool outputs; English text only.
- Notable 2.5: A vendor guard that compares itself to Jev and Bedrock Guardrails on a leaderboard.
- Novel 3: Builds on the vllm-sr d3-mini base.
Comments