Hacker News Reader

Comment by Akhiljp | Hacker News Reader

I beg to differ - manipulating a model to extract data is worse that protecting sensitive data. I am curious as to why this is not a big space yet?

View on HN

We already have many components and solutions for "firewalls", like everything we've done with microservice isolation. If you mean "sandbox" more than "firewall", there are 100s since Claude Code came out

So my first thesis that I want to prove is - are all enterprises going to start self hosting open source models ? If yes, then one will need to deploy a solution around the models to act as a firewall. A firewall fine tuned for the context that’s coming in and coming out of a model. Way different from how microservices work today.

1. Most enterprises are not self hosting at this point, hardware is too expensive. We want model choice. Eg. we are buying Fireworks.ai, however this is irrelevant to point 2

2. "firewall" is the wrong term, you are looking for "guardrails" if you are talking about prompt protections. Here again, there are 100s of projects and takes on this. Ai is the hottest space right now, people are thinking about all of these things. This piece is central to Fable access being shut off.

Ack on the guardrail analogy. On fireworks - do you just run serverless or do you take their dedicated instance. My question is how do you ensure that the model is not being manipulated to getting stuff it shouldn’t be.

"guardrail" is not an analogy, it is the term used in industry, you can search "ai guardrails" or "agent guardrails"

My bad. Wasn’t aware. Thank you.

Looking forward to understanding the second part of this.

In fireworks - how does one build guardrails or it does not matter to companies hosting their models on fireworks?

It doesn't matter where the models are, that's just a traditional API server. The agent harness is the thing that can read files and make http calls (to exfiltrate) and that's regular software running somewhere else. So one typically puts the harness in a sandbox and puts extra LLM calls in to guardrail (prompt / tool call checks).

I'm planning to do something like this for myself as an OpenCode plugin, but to stop my message and record it for clarity or block angry

Reply on news.ycombinator.com