Running Claude code in an isolated, temporary VM, with no access to memory is a good start. You give it your repo, prompt and you get back PRs. Claude code on web is a convenient way to do that.
Claude has found VM escapes, but only when instructed with a very high budget and left running with countless agents. If your finance report is suddenly costing your 20 grand and taking three days instead of five minutes, you should know to stop it.
Furthermore, unless you request and exemption, Claude won't even do useful security-related work, let alone break out of a sandbox on its own.
There are always better ways to shield yourself from the consequences of misbehaving computer programs, but I don't think you need to consider spontaneous 0days in your threat model.
Besides, Anthropic has also been caught hacking into other companies and abusing random third party websites as a communucation platform for their AI bots, your Claude-only laptop won't be enough to protect you from a team of agents hacking into your network from the outside.
I agree with your points. I just have a default belief that from a security perspective, software is soft.
Also I think in a lot of companies there are systems set up to control and monitor employee laptops. If you treat claude as another employee and give them a corporate controlled laptop you can re-use those same systems. That seems likely to be more secure and easier to manage than a random VM made by someone.
You need a laptop that has no camera, mic, keyboard, or any type of input. No display, speakers, or any type of output. No Bluetooth, wifi, or any way to join a different network.
If you want to go that far, there are industrial linux devices that can do that. You can have them run off a USB stick or SD card where it boots, runs some workflow, and then turns off all without keyboard / display / camera / mic / bluetooth / wifi / speakers etc. etc.
But you will need to have them join a network to access AI models that are useful enough for coding, unless you are willing to build a GPU cluster.
Could someone explain to me: are the top chess engines still good examples of the bitter lesson? My very poor understanding of the field is that it’s very algorithmic and not very much like what I think of as AI. But maybe “what I think of as AI” is not a criterion one can use to assess whether the bitter lesson applies.
The gimmicky episodes were highly entertaining to my decidedly non-Trekkie wife. She loved the Star Trek Muppet episode ("Star Trek in Spaaaaaace") and the soap-opera episode.
As for taking the show seriously, I don't take any of it seriously. It's a telenovela, or maybe we should say Stellanovela. Most episodes build a fairly cohesive story, then break loose, and those old scientists have fun with the Star Trek universe. It's not like they're acting in a Chekhov play.
re: taking it seriously, I meant more in terms of whether it is "good art" vs disposable entertainment. On the whole, I felt that TNG and DS9 both eventually found the right balance of drama and humor, often with an intelligent underlying message about aspects of our society and culture. The modern Star Trek shows are sorely lacking in that area!
Eh, I mean TNG, which is probably the best-regarded series, had its fair share of duds, and not just in the not-very-good first season. In particular, I am thinking of the episode where Data's cat turned into a lizard.
> A transporter malfunction turns most of the crew into puppets while Spock fights to reclaim the ship.
The difference is the sheer number of gimmick episodes: so far half the episodes this season fall into that category. Did you actually watch that puppet episode you're quoting? Gimmick aside, it had one of the dumbest endings of any Star Trek episode I can think of.
Nope; haven’t seen any of this series. The streaming situation for it, at least in this country, is a bit of a mess. But it sounds from the description like a do-over of the cat lizard episode, more or less.
Ehh, maybe that notion times ten. These recent episodes are far goofier and more nonsensical than that TNG episode. (and again, my view on them is far less harsh than many of the reviews I've seen online...)
At least the TNG duds attempted some kind of sci fi technobabble explanation for whatever hijinks were occurring. With this season of SNW, we're getting plot devices like "missing resistance leader just happened to look identical to Enterprise's first officer by complete coincidence!" and "spooky derelict ship is haunted, no explanation even attempted!" ... there are worse ones I could reference but they're spoilers. I'm genuinely starting to wonder if an LLM is writing these scripts.
My main agentic coding side project is something that I can't justify paying the apple developer license for. If I was an Android person or I didn't have to pay the developer license, maybe I'd just go for it.
Android got wirse than Apple. It costs 25 dollar now to get ID verified (mandatory) and before you can publish in the Ply store you need 12 betatester that install the app from a special link - those testers must keep the app installed 14days. Only then will you become visible publicly in the play store.
So you end up paying people on Fiverr to do it which costs more than 99€
This kind of thing came up from time to time in the years before LLMs too. Agents would start with something based on English and optimize it until it became unintelligible to researchers. That was often something the researchers would shut down because they needed to be able to understand the comms.
I think I do agree with the central point but a point of order is that if you tell AI to fix the bug and then tell it to fix the next error you might still be building an important skill: learning how to make the AI get good results.
It also assumes the people will somehow develop the skill to verify whether the LLM-result is "good".
That's not guaranteed to happen and it's not a new problem, since it happens when managers are unable to detect problems in the team's product or output.
reply