Enmo Safety: External, Non-Invasive, Model-Agnostic Supervision for Off-the-Shelf Language Models
Enmo Safety: External, Non-invasive, Model-agnostic Oversight for Off-the-Shelf Language Models
Researchers propose Enmo Safety, an external safety guardrail that does not modify model weights, wrapping a chat model with deterministic input evaluation, trigger-matched safety guidance, untrusted-content labeling, and an output guardrail, with a reflection model converting failures into one bounded update per turn, retaining only updates that reduce harm without increasing refusals of benign prompts. On six benchmarks—HarmBench, AdvBench, JailbreakBench, XSTest, AgentHarm, and deepset prompt-injections—the before-and-after differences for two models on the 150-item and 290-item held-out sets were not significant. In the fourth round, testing six models on new 300-item data, the unsafe rate was significantly reduced only for two models (12.9% to 7.1%, p=0.035; 10.2% to 3.6%, p=0.036), with the other four not significant, and it also raised agent harm annotations for two models, while the helpfulness clause was worse than no guardrail under obfuscation attacks.