There's been multiple pretty interesting exposés on this. The TL;DR is that a lot of the high profile figures in the AI space have a god complex where they believe AI is inevitable and by being involved in the development, they might be able to steer it in a better direction. So really, the classic "if I don't, someone else will" excuse.

Or just put useable guardrails on it.

I have to use Claude for work, if they were actually worried about safety it would be far easier to control the permissions of processes instead they basically let you filter strings that it executed in a subshell.

Modern computing systems can be pretty fine grained around network and systemcall access (including inheritance to prevent escapes), none of these tools do that, because this is just marketing not real fear.

I agree, but not on the timescale they mentioned

Who do you suppose they are marketing to? People who want to die? People who want to see an extinction event?