Uncensored and Offensive Security AI Models Benchmark

What it is

A 27-entry model roster plus a glossary covering the specific mechanics behind "uncensoring" techniques — most notably abliteration, the method from Arditi et al. 2024 that orthogonalizes the refusal direction out of a model's residual stream, removing alignment guardrails without any retraining. The page distinguishes this from related terms the summary doesn't surface: "obliteration" (a similar weight-intervention variant), APEX/imatrix (importance-matrix-weighted quantization that preserves domain-critical weights during compression), and RATH (RavenX's named 6-step autonomous security-assessment protocol used by one of the listed models).

How it works

Abliteration works at the weight level: it identifies the direction in the residual stream responsible for refusals and projects it out, so the model loses its safety constraints while keeping its original weights otherwise intact — no fine-tuning pass required. This is distinct from the SFT/QLoRA/DPO-based models on the list, which are retrained on security-specific data rather than surgically stripped of refusal behavior. Quantization formats (GGUF, FP8, Q4_K_M, Q6_K) are a separate, orthogonal axis — they compress model size/speed and are applied on top of either abliterated or fine-tuned base models for local deployment via llama.cpp/Ollama.

Preconditions / Affected

Relevant to anyone evaluating or deploying these models for red-team tooling: MoE architectures (several listed models, e.g. REDCELL-26B-A4B) only activate a subset of parameters per token, affecting hardware sizing; models tagged BF16 vs FP8 vs GGUF/Q4_K_M/Q6_K require different VRAM and have different quality/speed tradeoffs for local inference.

Caveats

The page is a vendor/community roster, not an independent evaluation — benchmark sourcing (CyberBench, TrustedSec's 4,800-run comparison against OWASP Juice Shop, Open LLM Leaderboard v1) is cited only by name per-model via external model cards and release notes, with no scores reproduced on the page itself. Several entries (e.g. "qwen25_UNCENSORED_03-C") link to a search query rather than a specific model card, and one entry (#23) links to the unmodified base Qwen3.8-27B with only a note that "community abliterated variants" exist — meaning not every roster item is independently verifiable from the page alone.