Metraipolis › archive › dc_d76074c375b5

RS-2026-080 - Capability or Compliance: What Abliterated Models and Guardrail Bypass Actually Reveal About AI Safety Alignment (2026) · OpenAlex W7214084369

Source: OpenAlex
Original: https://doi.org/10.5281/zenodo.22919815
License (as archived): CC-BY-4.0
Archived: 2026-10-08T04:18:31.880Z
SHA-256 of the archived text: 82590c1bd586d9e89f159e855f4dd7b07f42cbee583b3f9d723eef36429d60d6
Stable URL of this page: https://metraipolis.com/documento/dc_d76074c375b5

How to cite

RS-2026-080 - Capability or Compliance: What Abliterated Models and Guardrail Bypass Actually Reveal About AI Safety Alignment (2026) · OpenAlex W7214084369 — OpenAlex, https://doi.org/10.5281/zenodo.22919815 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_d76074c375b5 (Metraipolis archive dc_d76074c375b5, sha256 82590c1bd586d9e89f159e855f4dd7b07f42cbee583b3f9d723eef36429d60d6)

BibTeX

@misc{dcd76074c375b5,
  title = {RS-2026-080 - Capability or Compliance: What Abliterated Models and Guardrail Bypass Actually Reveal About AI Safety Alignment (2026) · OpenAlex W7214084369},
  howpublished = {OpenAlex},
  url = {https://doi.org/10.5281/zenodo.22919815},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_d76074c375b5; sha256 82590c1bd586d9e89f159e855f4dd7b07f42cbee583b3f9d723eef36429d60d6},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: OpenAlex, https://doi.org/10.5281/zenodo.22919815. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Opera scientifica in accesso aperto: scheda dai metadati OpenAlex (article).
Titolo: RS-2026-080 - Capability or Compliance: What Abliterated Models and Guardrail Bypass Actually Reveal About AI Safety Alignment
Autori: Richard Barron
Anno: 2026 · Sede: Zenodo (CERN European Organization for Nuclear Research)
DOI: https://doi.org/10.5281/zenodo.22919815
Accesso aperto: green · testo: https://doi.org/10.5281/zenodo.22919815 · licenza del testo: cc-by
Citazioni ricevute (OpenAlex): 0
Temi (OpenAlex): Adversarial Robustness in Machine Learning; Ethics and Social Impacts of AI; Topic Modeling

Abstract (testo degli autori, licenza cc-by):
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.

Fonte: https://openalex.org/W7214084369 — metadati OpenAlex (CC0 1.0), consultati il 2026-10-07 via api.openalex.org. Interrogazione: «AI safety alignment».