On November 2, OpenAI said it was preparing a new internal security testing sprint. The company plans to bring its own teams together with external researchers and security partners, using its models in the testing. Cathedral is a partner in the upcoming exercise. The announcement does not yet contain results from the sprint.
In such an exercise, a model could help reviewers look for recurring issues or construct test cases. A model's suggestion does not establish a vulnerability on its own. That requires reproducible observations, a clearly identified affected system and a check that the problem actually disappears after a fix.
A useful assessment of future results would distinguish initial signals, confirmed issues and deployed fixes. These numbers answer different questions. For example, more signals could reflect broader testing without by itself showing stronger defenses. The exercise can be assessed when concrete results are available.
*Cover: conceptual AI illustration.*
