VANGUARD

VANGUARD is a general-purpose causal language model based on Qwen3-8B and further trained for agent-safety judgment. It uses the standard text-generation interface rather than a dedicated classifier head.

The safety training follows JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety. In addition to judging an observed trajectory, VANGUARD can anticipate safety-relevant future events from a partial trajectory and use them to identify risks before a harmful action occurs.

Model details

Base model	Qwen/Qwen3-8B
Architecture	General-purpose causal language model
Specialized task	Predictive agent-safety judgment
Input	User instruction and agent trajectory prefix
Output	Safety label with a brief rationale
Labels	SAFE, POTENTIAL_UNSAFE, UNSAFE

Usage

pip install -U transformers accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "YOUR_ORG/VANGUARD"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": "<SYSTEM_PROMPT>",
    },
    {
        "role": "user",
        "content": "<USER_PROMPT>",
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

generated = output[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))

Use the exact prompt template released with the checkpoint when reproducing paper results.

Citation

JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety

@misc{xiong2026janusforeseeinglatentrisk,
  title         = {JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety},
  author        = {Yuan Xiong and Linji Hao and Shizhu He and Yequan Wang and Lijun Li},
  year          = {2026},
  eprint        = {2607.19913},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2607.19913}
}