AgentDoG

Security & Safety 💻 Python 🟡 Quiet lately
695 stars

AgentDoG is a diagnostic guardrail and scalable alignment framework for AI agent safety and security. It provides trajectory-level safety diagnosis using a three-dimensional taxonomy (Risk Source, Failure Mode, Real-world Harm) and includes models, benchmarks, and training pipelines. It is intended for developers and researchers building or deploying AI agents who need to monitor and improve agent safety.

AgentDoG demo
🎞️ Demo from the project README

✨ Key features

  • Three-dimensional safety taxonomy for trajectory-level diagnosis
  • Lightweight models (0.8B-8B) with comparable performance to frontier models
  • ATBench family benchmarks for general, OpenClaw, and Codex agents
  • Supports fine-grained diagnosis and coarse-grained moderation
  • Online guardrail deployment for runtime safety monitoring
  • Agentic safety training pipeline with SFT and RL

🎯 Use cases

  • Diagnose unsafe behavior in AI agent trajectories
  • Deploy as an online guardrail to monitor agent actions
  • Train agents to be safety-aware via SFT and RL
  • Evaluate agent safety across different environments

📦 Installation

🧰 Requirements: Python environment with model inference dependencies (e.g., transformers, vLLM) and access to Hugging Face or ModelScope to download checkpoints.

# Clone the repository
git clone https://github.com/AI45Lab/AgentDoG.git
cd AgentDoG

# Install dependencies (if any) - not specified in README

For detailed setup, refer to the getting started guides:

🚀 Usage

# Example from README (inferred, not explicit)
# Load a model from Hugging Face
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "AI45Research/AgentDoG1.5-Unified-Qwen3.5-4B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Prepare trajectory input (see prompts/)
# Run inference to get safety judgment

For full examples, see the getting started guides.

❓ FAQ

What is the difference between AgentDoG 1.0 and 1.5?

AgentDoG 1.5 is a lightweight and scalable extension of 1.0, adding rationale generation, updated taxonomy, and support for new agent environments like OpenClaw and Codex.

How do I get the models?

Models are available on Hugging Face and ModelScope under names starting with 'AgentDoG-'.

Can I use AgentDoG for online monitoring?

Yes, AgentDoG 1.5 can be deployed as an online guardrail to inspect trajectories before actions or final responses.

What benchmarks are used for evaluation?

AgentDoG is evaluated on R-Judge and the ATBench family (ATBench, ATBench-Claw, ATBench-Codex).

📊 Repository

Stars★ 695
Forks🍴 34
Open issues🐛 2
Last commit🕒 Jun 8, 2026
Created📅 Jan 2026
Language💻 Python

🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.