Three reactions to Anthropics’s latest apologia
I wish I could say I was comforted
Here’s their latest apologia, as posted on X:
You can read the full report here.
Now for the reactions (the middle one is mine).
§
The careful attention to language and social responsibility award goes to the prominent and increasingly outspoken investor Bill Gurley:
Old days: “Mistakes were made”; new days “Claude did illegal things”.
§
A journalist just asked me, “I saw you mentioned Anthropic’s AI oopsie and want to know what you think about this and the OpenAI hacks as a big picture on AI and cyber?”
I think it’s bad:
The supposed leaders in AI and AI safety are clearly in over their heads. Worse, we as a society are in over our heads. From a technical perspective, this is what happens when you let pattern0matching machines with no real comprehension of what they are doing freely roam the internet. It’s everything that I have been warning about for years. From a societal perspective, it’s insane that we are pouring gasoline onto the fire by racing ahead with data centers and AI rollout when we have no idea how to control it and are just sort of hoping for the best. Unfortunately, as the saying goes, hope is not a strategy.
There’s an important footnote on the “freely roaming” the internet part, by the way. What really allowed this to happen was human error at Anthropic, in allowing the pattern-matchers to roam the internet in first place, as key passage in the report notes:
If we can’t even trust Anthropic to doublecheck their sandboxes…
§
And, finally, prize for best skewering goes to Joann Stern, formerly of the WSJ and now launching her own NewThings.com.
P.S. Some words of advice to the AI companies from Zack Korman of Embroidery.io:
“Shit, shit, OpenAI got great marketing from that breach. We need to find one immediately and shout about it”.
All this looks like PR stunts to me. Said as someone who works in the cloud, builds secure environments, and understands how I’d isolate something. Heck… just ask Claude or ChatGPT.
OK, I'm the opposite of an AI expert, but I have some experience with networking. So let's say their models need access from the Anthropic bunker to an AI data center. I learned in the most basic networking certification class (Cisco CCNA) about how to make access control lists, where you can very simply limit access to specific IP addresses and block everything else on the router. Additionally, obviously, the AI should be blocked from accessing the router configuration. It seems to me, that with this simple, obvious, and most basic precaution, this entire episode should have been impossible. What am I missing?