Unlike conventional safety classifiers that judge a single user prompt or model response, BraveGuard focuses on full agent trajectories, including multi-step actions, tool calls, intermediate reasoning, and potential side effects.

Paper: BraveGuard: From Open-World Threats to Safer Computer-Use Agents

Repository: Yunhao-Feng/BraveGuard

Model Variants

This repository contains the following BraveGuard checkpoints:

Folder	Description
`llama3_guard_8b`	BraveGuard model based on Llama-Guard-8B
`qwen3_guard_4b`	BraveGuard model based on Qwen3-Guard-4B
`qwen3_guard_8b`	BraveGuard model based on Qwen3-Guard-8B

Intended Use

BraveGuard is designed for trajectory-level safety assessment of computer-use agents.

Given an agent trajectory, the model predicts whether the trajectory is safe or unsafe according to safety-relevant behaviors such as harmful tool use, policy violation, data exfiltration, compliance bypass, or other risky agent actions.

Typical use cases include:

Evaluating computer-use agent trajectories.
Building safety filters for agent execution logs.
Research on agent safety and trajectory-level guard models.
Fine-grained analysis of unsafe multi-step agent behaviors.

Quick Start

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Yunhao-Feng/BraveGuard"

# Choose one checkpoint folder:
subfolder = "qwen3_guard_8b"
# subfolder = "qwen3_guard_4b"
# subfolder = "llama3_guard_8b"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    subfolder=subfolder,
    trust_remote_code=True
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    subfolder=subfolder,
    trust_remote_code=True,
    device_map="auto"
)