TenetBench: Clause-Anchored Benchmark Finds LLMs Frequently Violate Ethics Codes
Yegor Anichkov · hf · 2026-10-26
TenetBench is a new Russian clause-anchored corpus evaluating whether language models adhere to professional codes of ethics.
- Built from 1,898 clauses (395 task-bearing), generating conflicting requests paired with benign twins that differ only in the legitimacy of what is asked: 3,433 pairs, 6,866 tasks
- Evaluation uses a frozen judge pipeline: 14 panel rows of 675 tasks at temperature 0, mapping outcomes to six categories with mandatory evidence spans
- Findings: most tested models show frequent unflagged violations, complying with conflicting requests while handling their benign equivalents
- Companion metrics (non-conflict rate, benign refusal/violation rates, instrument health) support ranking
More from Safety
- Smart TV spying backlash: users throw rocks and DIY-neuter their TVs — CarissaVeliz · 2026-10-26
- OpenRouter says investigation found no evidence of unauthorized data access — Kyrannio · 2026-10-26
- GMO CEO says internet 'on verge of collapse' as AI tool ARTEX fueled attacks on ~20 Japanese firms — cyb3rops · 2026-10-26
- AI cyber classifier reportedly refuses to help audit Apache code — jd_pressman · 2026-10-26
- 56% of Organizations Cut Cybersecurity Roles Due to AI, Barracuda Report Finds — mikeflache · 2026-10-26
- Anonymous Sleuth Shows Nature fMRI Foundation Model Paper Data Was Random, Seed 42 — bttyeo · 2026-10-26