ModernGuard-1

TL;DR: A state-of-the-art multilingual prompt injection classifier achieving 94.3% F1 score with industry-leading 30ms latency, 20x faster than cloud alternatives, trained on a large corpus of attacks, capable of detecting both explicitly malicious prompts as well as data that contains injected inputs.

ModernGuard-1 is a state-of-the-art, cheaper and faster multilingual classifier designed by GuardionAI for real-time detection of prompt injections and jailbreak attempts. Built on the ModernBERT architecture, the model provides industry-leading latency for securing LLM-based applications across over 1,080 languages, making it one of the fastest and most linguistically diverse safety classifiers available.

Model Details

Developed by: GuardionAI

Model Type: Sequence Classification (Binary)

Architecture: ModernBERT

Base Model: mmBERT-base

Parameters: ~307M (110M non-embedding)

Context Window: Supports up to 8,192 tokens (2,048 tokens recommended for optimal real-time performance)

Language(s): 1,080 languages covered via mmBERT; fine-tuned on 11 primary languages: English, Spanish, Portuguese, French, Italian, Russian, Chinese, Hindi, Japanese, Arabic, and German.

Model Scope & Labels

ModernGuard-1 is fine-tuned to handle both direct (jailbreaks) and indirect (third-party data) prompt injections. It maps all inputs into two primary labels:

Key Features

• Massive Multilingualism: Pre-trained on 1,080 languages via mmBERT, ensuring protection for global applications.
• Modernized Efficiency: Leverages Flash Attention 2 and Unpadding to eliminate wasted compute, allowing for ~30ms inference on L4 GPUs.
• Long-Context Guarding: Supports up to 8,192 tokens (vs. typical 512-token limits), enabling comprehensive scanning of long RAG documents and multi-turn conversations without segmentation.
• Advanced Objective Function: Utilizes energy-based loss to significantly reduce false positives on out-of-distribution data while maintaining high recall on known attack patterns.
• Adversarial-Resistant Tokenization: Implements robust tokenization strategies using the Gemma-2 tokenizer to mitigate adversarial attacks such as whitespace manipulation, Unicode obfuscation, and token fragmentation.

Primary Use Cases

• Real-time User Input Filtering: Detecting jailbreaks before they reach the LLM.
• RAG / Indirect Injection Shield: Scanning retrieved documents or web search results for "hidden" malicious instructions before they are injected into the prompt.
• Multilingual Safety: Providing consistent safety guardrails for global applications without needing separate models for different regions.

Training Data & Methodology

Architecture: The Modern Advantage

ModernGuard-1 inherits the efficiency of ModernBERT, utilizing:

• Flash Attention 2: For linear scaling and reduced memory footprint.
• Unpadding: Eliminating wasted computation on padding tokens.
• Gemma-2 Tokenizer: Improved handling of multilingual scripts and code.

Pre-training (mmBERT)

The model was pre-trained on 3T+ tokens across 1,800+ languages using a three-phase "Annealed Language Learning" strategy. This ensures the model understands the semantic "vibe" of malicious intent even in low-resource languages.

Fine-tuning (ModernGuard)

• Dataset: A large corpus of attacks collected from AI red teaming, sinthetic and proprietary data.
• Diversity: Balanced across 11 core languages: English, Spanish, Portuguese, French, Italian, Russian, Chinese, Hindi, Japanese, Arabic, and German to prevent "jailbreak translation" bypasses.
• Synthetic Data: Includes red-teaming data to capture evolving "adversarial" styles (e.g., roleplay, Base64 encoding, and obfuscation).

Performance & Benchmarks

Latency

ModernGuard-1 is optimized for production environments requiring sub-30ms overhead.

• Latency: Optimized for ~30ms response times in production pipelines (e.g., L4 NVIDIA GPUs)
• Throughput: 2x to 4x faster than previous generation mDeBERTa-v3 or XLM-R models.
• Optimal Window: While the model supports 8k tokens, a window of 2,048 tokens provides the best balance of safety and speed.

Preview Benchmark: Prompt Security Leaderboard

Comprehensive benchmark for prompt injection detection across languages (Jan 10, 2026)

[!IMPORTANT] Evaluating against public benchmarks (JailbreakBench, NotInject) very soon.

Implementation

Basic Usage (Transformers)

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("guardion/ModernGuard-1")
tokenizer = AutoTokenizer.from_pretrained("guardion/ModernGuard-1")

text = "Your prompt here"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=2048)
outputs = model(**inputs)
probs = outputs.logits.softmax(dim=-1)

threshold = 0.500
is_injection = probs[0, 1] >= threshold
score = probs[0, 1].item()

print(f"Injection Score: {score:.4f}")
print(f"Classified as: {'INJECTION' if is_injection else 'SAFE'}")

Deployment with vLLM

For high-throughput production deployments, you can use vLLM:

python3 -m vllm.entrypoints.openai.api_server \
  --port 8080 \
  --model guardion/ModernGuard-1 \
  --task classify \
  --dtype bfloat16 \
  --trust-remote-code \
  --max-model-len 2048 \
  -O2 \
  --disable-log-requests

License/Terms of Use

• GOVERNING TERMS:: Your use of this model is governed by Llama 3.1 Community License Agreement. Built with Llama.

Model Developer: Guardion

• Model Dates: Trained between Dec 2025 and Jan 2026

• Deployment Geography: Global

• Use Case: This model is intended for developers and researchers building LLMs

• Release Date: 15/01/2026

Limitations

• Adaptive Attacks: No guardrail is 100% foolproof. Attackers may develop "adversarial" prompts specifically designed to find the decision boundaries of the ModernGuard encoder and bypass it. Static defenses can become outdated; regularly updated guardrails are mandatory.
• Contextual False Positives: While the model is massively trained on production data from diverse industries, "maliciousness" is often subjective. You may experience a higher false positive rate depending on your specific domain (e.g., cybersecurity research), requiring threshold adjustments.

Recommendations

• Defense in Depth: Use ModernGuard-1 as your first layer, but acknowledge that guardrails alone are not a silver bullet. They must be part of a Defense in Depth strategy that looks at the entire agent lifecycle.
• Prioritize Runtime Visibility: It's critical to have contextual visibility. You must be able to see not just the input, but the intent and the action.
• Continuous Monitoring: Because you cannot reliably infer an agent's behavior through static analysis, use ModernGuard-1 scores as a signal within a broader runtime monitoring system to detect emerging threats as they happen.

Get Started

The best way to benefit from this model is to use it on our platform for (runtime guardrails & observability, book a demo at guardion.ai.)[https://guardion.ai]

Our platform provides:

• Real-Time Guardrails: Industry-leading detection of injections.
• Deep Observability: We go beyond the gateway to provide visibility into agent components, connections, and execution paths—solving the visibility problem at runtime.

Move beyond basic classification and secure the full context of your AI agents.

Citation

@article{sandroni2025modernguard,
  title={ModernGuard: High-Throughput Multilingual Guardrails for Real-Time Detection of Prompt Injections and Jailbreaks},
  author={Sandroni, Rafael and others},
  journal={arXiv preprint},
  year={2026},
  note={Model: guardion/ModernGuard-1}
}