ATechGuy

A few startups [1,2] also offer infra for private AI based on confidential computing from Nvidia and Intel/AMD.

1. https://tinfoil.sh 2. https://www.privatemode.ai

Maybe the author can clarify.

Sounds promising. How do you propose we create "hardened" projects?

This is scary!

Wondering if used tried stopping such bots with Captcha?

TL;DR Yes, because the current AI has the same last-mile problem as the tech around self-driving cars.

> I was trying to create a stack / undo context for REPLs that LLMs could use, since they often put themselves into bad states, and reverting to previous states seemed useful

This is interesting! How did you end up achieving this? What tools are available for rolling back LLMs doing?

Could you use ngrok and dropbox to put together something like this quickly?

> Security team is just another compliance department. It does not bring in money.

This. Thanks for highlight it!

> Security is always optional.

Deeply buried, but highly insightful. This is what Cybersecurity startup founders and professionals must pay attention to before they commit to the field.

Yes, that's the state of the art unfortunately. There are a number of vendors who can scan your source code and generate a list of FOSS libraries with licensing info (SBOMs). If we can find a way to get this money flowing to our fellow FOSS devs, that would be a huge win for the community.

What does it offer compared to ChatGPT?

And you think nobody (professional hackers?) can put together a "virtual TPM" that falsifies real hardware info? I think there are much simpler solutions, but the big tech wants to retain the control.

I don't think tracking everything is the way to go; info would get outdated very soon and tracking compromises user privacy. A simple solution could be to throw a challenge that humans can easily solve, but agents absolutely cannot now or in the future (think non-audio/visual/text).

Looks like detecting real humans apart from agents is going to be an arms race if the detection is based on browser/device fingerprinting or visual/audio captchas; AI will only get better.

What are captcha alternatives that can block resource consumption by bots?

The choice is simple: work for next Nvidia or pursue academia?

That's exactly why I've built a tool using LLMs that filters out this garbage before showing me the content.

YC has listed some ideas: https://www.ycombinator.com/rfs

> Dev containers should do this, but I don't like the performance What kind of dev container have you tried? What is the source of perf overhead?

This. I'm surprised that even a lot of smart and tech savvy folks are not able to see these pitfalls.

> There are some startups building capabilities apis to limit that but most websites/apps either don’t have the resources or aren’t willing to expose those capabilities.

Care to elaborate with example(s) of a startups doing this?

One idea maybe to submit to an industry conference and review the feedback. Usually they only require you to submit a short abstract and the turnaround time is a few days.

I wonder if browsers have a future.

I think you are assuming that inference happens on the same machine/VM that executes code generated by an AI agent.

I mean performance overheads of an OS process running in a VM to (vs no VM) and additional resource requirements for running a VM, including memory and additional kernel. You can pull relevant numbers from academic papers.

VM is not the right abstraction because of performance and resource requirements. VMs are used because nothing exists that provides same or better isolation. Using a throwaway VM for each AI agent would be highly inefficient (think wasted compute and other resources, which is the opposite of what DeepSeek exemplified).

For all speculation around supply chain attacks with fake Github stars, the article says:

"our study does not find any evidence of fake stars being used for social engineering attacks"

Care to elaborate?

Congrats on launching!

> spin up hundreds of browser sessions in secure, isolated environments, with sub-second launch times

What tech do you use for this, and where is this hosted?

> Modal has an isolated container runtime that lets us share each host’s CPU and memory between workloads.

Looks like Modal hosts workloads in Containers, not VMs. How do you enforce secure isolation with this design? A single kernel vulnerability could lead to remote execution on the host, impacting all workloads . Am I missing anything?