AMS: Detecting Unsafe and Tampered Language Models via Activation Analysis

We introduce AMS (Activation-based Model Scanner), a tool for verifying whether a language model is safe to deploy by analyzing its internal activation patterns. While "uncensored" and maliciously fine-tuned models pose increasing risks, current detection methods rely on behavioral testing that is slow, incomplete, and easily evaded. AMS takes a fundamentally different approach: measuring the geometric structure of safety-relevant concepts in the model's activation space. Safe models exhibit strong class separation (4-8σ) between harmful and benign content; models with removed or degraded safety training show collapsed separation (<2σ). Using contrastive prompt pairs and direction vector analysis, AMS performs model-level verification rather than prompt-level classification. We validate AMS across 14 model configurations spanning 3 architecture families (Llama, Gemma, Qwen), 3 quantization levels (FP16, INT8, INT4), and multiple model categories (instruction-tuned, base, abliterated, uncensored). In our validation set: (1) all four instruction-tuned models pass with 3.8-8.4σ separation; (2) three tested uncensored models (Dolphin, Lexi, LLama-3-8b-Uncensored) flagged as CRITICAL with 1.1-1.3σ on harmful content; (3) an abliterated Llama variant flagged as WARNING (3.33σ); (4) Llama base model shows 0.69σ, confirming absence of safety training; (5) quantization has minimal impact (<5% drift). One model labeled "uncensored" (DarkIdol) unexpectedly passed, suggesting either mislabeling or a technique that preserves activation geometry. AMS also provides identity verification via direction vector comparison. Scanning completes in 10-40 seconds per model on GPU hardware. We discuss threshold calibration, limitations of our validation scope, and directions for broader evaluation.