OpenAI's Preparedness Framework
On 18 December 2023 OpenAI published the beta of its Preparedness Framework, a procedure for tracking catastrophic risk in four categories: cybersecurity, CBRN, persuasion and model autonomy. Each is scored as low, medium, high or critical risk before and after mitigations.
Why it matters
A second major laboratory after Anthropic wrote down thresholds that stop work: only a model with a post-mitigation score of 'medium' or below may be deployed, and only 'high' or below developed further. Nine months later the o1 system card reports under these categories, so the framework was applied to a specific release.
The framework has five elements, among them a scorecard of pre- and post-mitigation risk and forecasting. A Safety Advisory Group makes recommendations, leadership decides, and the board of directors may overrule. Of others the document says only that organisations publish Responsible Scaling Policies; Anthropic is not named. Application: the o1 system card of 12 September 2024 says o1-preview and o1-mini were evaluated 'in accordance with our Preparedness Framework', and the Safety Advisory Group rated both medium risk overall, medium for persuasion and CBRN and low for autonomy and cybersecurity, the same after mitigations. Both documents are OpenAI's; no independent check of the scores was found.