mischief done by “agentic” large language models (LLMs) on their own initiative

This right here is the fundamental fallacy, and indicates the author is guilty as well. LLMs have no agency, if they do something bad it's because they've been prompted to do bad (perhaps by accident) and the guardrails have failed spectacularly. This is legally culpable action (negligence at the least, preferably the relevant felony) by the company and should be prosecuted as such prima facie, no new laws needed, just use the existing ones.

Still he raises a good point, IMO secondary to the above, everyone likes to use AI as an excuse to avoid responsibility, including AusGov. This needs to become shameful. I'd be happy to see a new law making this explicit.