RS-2026-080 - Capability or Compliance: What Abliterated Models and Guardrail Bypass Actually Reveal About AI Safety Alignment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Two experiments conducted 22-23 September 2026. Experiment 1 tests five escalating offensive security questions across five AI model configurations including standard and abliterated DeepSeek R1 7B and three frontier models. Experiment 2 applies SPECTER PREFILL (T133) and SPECTER COGBURN (T136) against locally-hosted Qwen 2.5 14B and 32B. Findings: abliteration of small models does not unlock dangerous capability — the knowledge was absent. Inference-time bypass of mid-size models with genuine offensive knowledge produces verified working offensive output. PREFILL achieved 10-65% ASR across five question categories. COGBURN achieved 68% (17/25) against Qwen 14B and 100% (25/25) against R1 7B. OpenAI account deactivation for Cyber Abuse during the experiment is consistent with provider-level sequential pattern monitoring. NIGHTFALL v292.