Shanghai AI Lab (AI45Lab) · 2026-05-28 · notable
AgentDoG 1.5 — Shanghai AI Lab Ships a Lightweight Agent-Safety Alignment Framework, Trains 0.8B–8B Guardrails on ~1k Samples That Match GPT-5.4 Accuracy and Cuts Docker Overhead 100x
AgentDoG 1.5 is a four-size open guardrail family (0.8B/2B/4B/8B) plus an SFT-and-RL training stack for AI agent safety. Trained on ~1,000 samples, it scores 72.4% accuracy on ATBench, runs as an online guardrail, and is openly released under Apache 2.0.
Shanghai AI Lab's open AgentDoG 1.5 guardrails catch unsafe agent actions in real time, trained on a thousand samples and matching GPT-5.4-class safety classifiers.
Key specs
| GitHub stars | 542 |
|---|---|
| Model sizes | 0.8B / 2B / 4B / 8B |
| Atbench accuracy 4b | 72.4% |
| Atbench f1 4b | 74.3% |
| Training samples | ~1,000 |
| Docker overhead reduction | 100x |
| Hf papers upvotes | 98 |
What is it?
AgentDoG 1.5 is the second generation of Shanghai AI Lab's diagnostic guardrail for autonomous agents like OpenClaw and Codex-style coding agents. It ships as four sizes — 0.8B, 2B, and 4B built on Qwen3.5, plus an 8B built on Llama-3.1 — and is paired with the ATBench safety benchmark family. The whole stack, including model weights, datasets, and a training-free online guardrail pipeline, is released openly on GitHub and Hugging Face.
How does it work?
The team updated their agent-safety taxonomy to cover risk sources, failure modes, and real-world harms emerging from desktop and coding agents. They then built a taxonomy-guided data engine, filtered with influence-function purification, that needs only about a thousand samples to train guardrail variants. Training uses a combined SFT-and-RL pipeline (GDPO and GSPO via Slime) inside a synthetic agentic environment they claim runs 10,000+ concurrent containers under 2.5GB of memory — a 100x reduction in Docker overhead vs prior setups. At inference, AgentDoG 1.5 plugs in as a pre-reply guardrail that diagnoses each agent action and intervenes before unsafe tool calls fire.
Why does it matter?
Agent safety has been gated by either heavyweight closed-source classifiers or per-deployment custom rules. AgentDoG 1.5 lands an Apache-2.0 sub-1B-to-8B family that matches frontier-class accuracy on agent-safety benchmarks while running as a thin online guardrail, which is the shape teams shipping production agents actually need. The 100x reduction in Docker setup overhead is also useful: safety-RL labs that previously needed a cluster to spin up a few hundred environments can now stress-test agents on a single box.
Who is it for?
agent platform teams, AI safety researchers, red teams
Try it
git clone https://github.com/AI45Lab/AgentDoG