Existing LLM safety measures operate as black boxes, lacking categorical precision and calibrated severity assessment. We propose **PRISM** (**P**rototypical **R**epresentation for **I**nternal **S**afety **M**apping), a fine-grained two-stage safety auditing framework grounded in internal mechanistic representations. Stage I localizes functional safety zones via spectral analysis and maps internal activations onto concept prototypes for camouflage-resilient categorical risk identification. Stage II decomposes violation severity across orthogonal dimensions through structured CoT reasoning, producing calibrated human-auditable ordinal verdicts. PRISM achieves a new best F1 of 77.60% on the adversarial ToxicChat benchmark and approaches 671B frontier judges within a 2.3% margin on severity calibration, demonstrating that mechanistic grounding and principled task formulation can substitute for scale in fine-grained safety governance.