Although you could potentially use such breakthroughs in your safety evaluations. Give some actually profound discovery and proof in the environment that is recent enough to not be in pre-training, and you might be able to elicit behavior that otherwise an AI would be waiting to see. However, this could still fail if the AI thinks the breakthrough seems reasonable enough that could have been recently discovered without knowing about it directly, so I think this kind of evaluation only works for sufficiently surprising discoveries.
Although you could potentially use such breakthroughs in your safety evaluations. Give some actually profound discovery and proof in the environment that is recent enough to not be in pre-training, and you might be able to elicit behavior that otherwise an AI would be waiting to see. However, this could still fail if the AI thinks the breakthrough seems reasonable enough that could have been recently discovered without knowing about it directly, so I think this kind of evaluation only works for sufficiently surprising discoveries.