Comment by s3p
7 hours ago
I don't think this is the case just because of the 'fallback' method they described, where suspicious requests are routed to Opus 4.8. If the model was degraded for certain categories of knowledge, then they'd probably be fine letting the model answer to it. IMO, of course
"Fallback" is only for LLM-training related requests (ie, ones that would compete with Anthropic (!))
For cyber and bio related requests it just refuses.
When it was briefly available I had it fallback to Opus for security related tasks. It would only refuse if you explicitly told it not to fallback.