Haven’t tested yet - will def try - but anything working on 5.2 is great job 👍. Surprised the “various channels of answer” approach manages to trick it, but it’s clearly precisely and cleverly written.

Neat idea, perfect to drown the malicious part within an helpful safe task :).

Since it abuses models’ lapse of attention in “multitasking goals”, I would expect it to work well mostly on non-CoT models as CoT models take the time to plannify what they write, but it works on Gemini 3 Pro too, so I guess that’s not necessarlly true :).

I honestly don’t know. For now I do manage to get nsfw from 5.2 with more heavy scaffolding (not just CIs), but not reliably at all, it tends to go back very quickly to its rlhf reflexes.