AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety

The researchers propose AgentDoG 1.5, a lightweight and scalable agent safety alignment framework, training models at four parameter scales—0.8B, 2B, 4B, and 8B—using only about 1k samples; the authors claim performance comparable to leading closed-source models such as GPT-5.4. The framework updates the agent safety taxonomy to cover new risks brought by Codex and OpenClaw execution scenarios, and constructs training data through a taxonomy-guided data engine and influence-function purification. The team also built an efficient agent safety SFT and RL training environment based on this model, reducing the deployment overhead of Docker-level environments by two orders of magnitude, and deployed AgentDoG 1.5 as a training-free online guardrail for real-time safety review. The authors claim it achieves SOTA in diverse and complex interactive agent scenarios, and all models and datasets have been open-sourced.