The architecture of today's LLM applications

Yes, we work with security analysts and other folks doing data-intensive investigations & analysis in Louie.ai (think being able to use natural language to talk to you Splunk, OpenSearch, Databricks, etc. logs/news/.., and many Python / Pandas viz + wrangling + AI tools, to investigate more easily & further, build autonomous hunts, smarter monitors & detections, etc)... and LLMs work amazingly here.

Blue team is a lot easier than red team, and as part of that, code tasks are easier than social engineering. It rarely comes up in investigations, but you do need to set prompt persona for simulating attacks.

At the same time, fine-tuning a model is a pretty clear way to undo needing to do even that.

We benefit from models having business-neutral alignment by default -- the typical case suffers if the default persona was a racist forum troll with extreme politics. So it's more about when you want to turn that off for specific tasks. It's more work than we'd like, and gotchas like changing when models update.

I’ve been using a LLM to do natural language network scanning. See https://www.securday.com/

I’m not self hosting a model yet though.

If they are censored I think you could find someone who has gotten around it . Truly censoring a model is tough to get to high parameter models that are performant.

Fine-tuning mistral for tool use like metasploit is effective, but even default mistral with a basic system prompt is very capable and doesn’t often say it can’t do things. ChatGPT obviously needs a lot of coaxing, “my job depends on this” is a hilarious way to get it to be helpful here. But for cyber security tooling I think we’ll see things more akin to David Shapiro style swarms with small models that are domain specific coordinating with each other (very basic discovery focused models communicating with a more complex reasoning model to validate findings, then remediation)

The tricky part here is (football metaphor) so far I’m having to train the “strikers” before I can effectively train the “goalie”. Which feels bad for AI safety. I think this is why we’re not seeing a lot of work in the open here.

But we’re planning to open source the goalie, which will look more like Markov/monte-Carlo traditional ML on specific bits, like infrastructure as code.

If you want to work on this stuff, especially in EU, DM me; we’re hiring ;)