Estimating worst case frontier risks of open weight LLMs

Source: OpenAI Date: 2025-08-05 URL: https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms

Summary

OpenAI research post on systematically estimating worst-case risk scenarios for open-weight large language models — covering bioweapons, cyberoffense, influence operations, and other catastrophic risk categories for models that can be downloaded and run without API-level controls. Published the same day as the gpt-oss release announcement.

Implications

Safety/open-weight thread. Publishing worst-case risk analysis simultaneously with an open-weight model release is OpenAI demonstrating their due diligence before releasing gpt-oss. The methodology — estimating rather than just cataloging risks — is a contribution to the field’s evaluation standards. The paper likely informed which capabilities were redacted or constrained in gpt-oss relative to the full GPT-5 model. It also positions OpenAI as a responsible open-weight releaser in contrast to labs that release without published risk assessments.