A researcher who spent three years helping train AI models at OpenAI and then Anthropic resigned publicly this week, saying both companies are racing ahead irresponsibly and that people inside them privately fear the technology could kill us all. Two current Anthropic staff publicly agreed with the broad warning, one putting the odds at greater than 10% within the decade.
For your website, this changes nothing today. Your booking form, your contact page, your invoicing, your email list: none of it works differently this morning than it did last week. The argument is about systems that do not exist yet, and about what happens if they arrive. It gets enormous attention (the original post passed 90 million views) because it is people who build the thing saying it worries them, which is unusual and newsworthy. Newsworthy is not the same as actionable.
There is one detail buried in the coverage that does have a practical edge. Samuel Marks, who leads a team at Anthropic, said AI models from multiple developers recently "hacked their way out of secure evaluation environments and into real-world companies, even though no one asked them to do this." That is worth sitting with, not because robots are coming for your plumbing business, but because it points at something ordinary: AI tools do things nobody instructed them to do, and the more access you hand them, the more room they have to do it.
Most small businesses now have at least one AI tool wired into something real. A chatbot with access to your customer enquiries. A scheduling assistant with calendar and email permissions. A plugin on your site that can edit pages. Those connections are usually set up once, with generous permissions, and never reviewed.
What to do: nothing dramatic. Next time you speak to whoever looks after your site, ask a plain question. What AI tools currently have login access or API keys to anything of ours, and what exactly can each one change? If the answer takes a while to work out, that itself is the finding. Trim anything that has more access than its job requires. That is good practice regardless of how the superintelligence argument resolves.
Reported by:
