NUWA RESEARCH RECORD
Findings from a pilot Anthropic–OpenAI alignment evaluation exercise: OpenAI Safety Tests
Original publication ↗ · BibTeX · RISThe introductory summary reports task- and reasoning-dependent results and notes that grader errors affected one jailbreak comparison.
Research question
What failures appear when another vendor runs its own safety tests?
Design & assumptions
OpenAI tested Claude and its own models on instruction hierarchy, jailbreaks, hallucination and scheming, with some external safeguards relaxed.
Findings
The introductory summary reports task- and reasoning-dependent results and notes that grader errors affected one jailbreak comparison.
Limits & counterevidence
The authors explicitly reject exact apples-to-apples comparisons and deployment-frequency interpretations.
Source evidence
No structured numerical extraction is available in this reading package.
Reading scope & review
Web-tool rendered extract captured through line399; introduction and summary selected. Detailed later grader-error section not captured/read in this extraction; linked full evaluation appendix not audited.
Model-assisted reading; independent evidence review pending.
Check the original source ↗