Forward AI Demo – Risk Mitigation and Security with Forward Networks

This is, uh, around, uh, risk mitigation and, uh, uh, and, and, uh, compensation control. So we are in the security team now and, uh, let's, let's jump to forward ai. I'm in the security team and I am dealing with a situation, and this is what the situation looks like in plain English.

I have a host badge zero one that has critical vulnerabilities. And, uh, the question I want to answer is, what firewalls do I have to block in order to remove access to the internet? Let me give you a little bit context of what's what we're trying to do here.

We have detected that there is a host, uh, it's called Badge zero one that has been detected to have, uh, vulnerabilities and patches are not available for that host yet, right? So I, but I still need to make sure that we are mitigating that risk, right? And one of the things that I need to do is block it, access to the internet so that it cannot establish a command control channel, uh, with, with, with some server out there.

Are you leading the witness a little bit by asking about the firewalls rather than just saying, how do I block this? Yeah, I could potentially, I could potentially ask like, how do I block access to this as well? Right?

But again, like I'm, I, I want like something more, I'm giving it a a bit more specific prompt. Yep. And again, like this is how it works in ai.

Like the more specific you are, the more likely that you'll get a direct answer to exactly what you want. Right? Right.

Uh, let's, let's, uh, hit this running. And while this is working, let's, let's look back at the traditional world and try to figure out, try to understand like how I would do this. Again, I would do something very similar, like the last use case.

I would do this hop by hop analysis to understand what is the path that this traffic would take through the network. What are the firewalls along the path that I would need to provision to block access? Uh, and now one like that task again, is time consuming, can easily take hours.

But what makes this problem even worse is it is entirely possible that this is not the only host that I have detected to be vulnerable. I could have a large number of these hosts and I need to do this activity for every single one of them, because that like spread around around the, around the network. Right?

And also, like what's, what's even worse about this is if I get this wrong, then I have, I would've just left the network open to attacks. Like, so I may not even resolve the issue and I may not even know about that, but I just asked forward AI to do this and I had it, uh, produce a plan, uh, and, and get details. So it did something very similar.

Like last time it got more details about this host badge zero one because, because it needs that context around what risk vulnerabilities are present. And then it did a path trace from that IP address of that host to a very broad range of IP addresses 0, 0 0 slash zero. Like it wants to know like exactly where all traffic can go from this particular host.

And based on the analysis, it, uh, it comes back with this answer, this badge zero one host reaches the internet through, uh, one on-premises firewall. Uh, it's SJC building one FW zero one, uh, which currently permits the egress traffic. Like that's the firewall that is allowing traffic to go out to the internet, right?

But that's what, that's what forward AI is saying. Mm-hmm. But again, like a good security engineer, I shouldn't just take, take, take this, take its word, right?

I need to validate this for myself. Mm-hmm. And like before, it is backed with evidence.

And here's the, here's here's what our day I saw based on which it made that as a, uh, assertion. Let's, uh, click, uh, I'm gonna click on the link at the bottom to look at this in more details. There are 98 different parts that traffic from this host can take to various addresses in the 0 0 0 slash zero.

And here's one part that I'm seeing, and this is the SJC building one firewall that is permitting the traffic to go through, right? But I can also use this as a starting point to dig in deeper and again, explore more, uh, more variance of this just to be sure that this is, this is the firewall that I really need to block. I'm gonna make one example change here.

I'm gonna change this broad 0 0 0 slash zero to one of the addresses on the internet. 8, 8, 8 8. That's Google DNS.

And then I see the one path that traffic takes from, from that IP address for 8, 8, 8 8, and again goes through this particular firewall. SJC building one, FW zero one. And why is this traffic going through this firewall?

Well, there is an access control, uh, routed nets interlink, that's the name of the access control. That's a, that's a rule that is present on this firewall that is permitting traffic, all traffic from that source IP address subnet. This is why we know traffic is egressing out to the outside world.

Mm-hmm. Now, uh, like before, uh, okay, I found I've detected the firewall that is permitting the traffic to go out. Let's have, uh, forward ai, if we collapse this section, I'm gonna ask forward AI to suggest firewall rules to block access.

Right. While it's doing that, there is a big gap between 8, 8, 8, 8 and 0, 0, 0 0. Yeah.

Um, I have 97 other paths. Should I spot check a few other? How do I know like that's the one for most?

Excellent. Excellent question. Excellent question.

So like, again, a more thorough analysis of this would, would have me go through, comb through those parts and make sure that there, that one particular firewall is, is the one that is allowing all of that traffic to go through that. That's the one common firewall. Is, is there a way to automate that or, or summarize that?

Like seeing 98 paths is great. Yeah. I mean, like one Of, one of the aside.

Yeah. One of the, uh, future items that we have for four days is to be able to automatically comb through all of those parts and, and actually give that assertion. Excellent.

Okay. Thank you. Wouldn't, wouldn't you be able to ask it to, uh, configure firewalls for all possible paths to 0, 0, 0.

Yeah. As one way, one way of getting through this, and it might all be the same firewall, but it may not. Absolutely.

Yeah, absolutely. Like, I mean we, yeah. So one of the capabilities that, uh, we will be adding to forward AI Is a traction.

Yeah. Okay. I gotcha.

And so here's, here's the CRI commands that it is, uh, suggesting. Again, it is, uh, what it's doing here again, like I'm seeing this, these commands, uh, live here with you. It is removing that routed nets to internet access list.

And it's updating this with one, like a top level rule that denies traffic from that host to any target. And it's, I think it's adding all of the remaining rules that were already present. So this, this is, uh, again, this needs to be validated, uh, before we execute, but like it's giving us a head start.

It's identified how the traffic flows from, from this vulnerable host out to the internet, identify the firewall, and it's giving us a head start on how to block access. Quick question. Did, did the prompt make the assumption you were referring to egress traffic from the host though?

That's, that's what I mm-hmm. Like could you go back to the prompt real quick? Yes.

'cause I'm curious, um, in which direction and how did it determine that? You're saying from the host of the internet Yes. Access to the Internet.

Okay, gotcha. From the, from the, from the, to the internet. So it assumed be grace up.

Exactly. Okay, gotcha. Marion, I have a quick question.

In the example you showed where it went across several countries, what policy takes precedent? Can I set a local policy based on local, um, regulatory requirements? Which would it take precedence over what's being typed there in the ui?

Can you elaborate a little bit more on what do you mean by local? So For like cross border data management, uh, EU act where I have to keep certain data in certain places, uh, and if I have data crossing, uh, private sensitive information, banking data has to stay in one place, but my network, I may have that same infrastructure just in a different country. So what policy takes precedent, what I've put in there or the network configuration?

Yeah. Uh, excellent question. So we have, we have the ability to bring in business data, business policies into the mix.

And, uh, and if and when we do that forward, AI will be able to take that into account as well, uh, when it makes suggestions. Is that a future state or that's current? It is.

Uh, we have the ability to pull in some, some of the external data, but we have not integrated that with forward AI yet. So that's, that's, thank you. Thank you.

Hey, Nikhil, quick question. Um, we, we kinda had a little in internal, uh, string going here. When, when you pull information on in the life devices security vulnerabilities, that's not trained data, right?

You're just going out to vendors and, and getting that information real time? Exactly. Okay.

We have, we have a database of all of the end of life, uh, operating system information, and then we pull that in and make that a part of the snapshot. Okay. How Frequently is that updated?

Just whenever you create a snapshot or on a regular cadence when you update whatever label I have to get back to you on the exact frequency of it. Uh, I don't know if it is live updated or it's updated with every release. I am like, not certain on the specific, but at least as frequently as every release of our platform.

And, and usually when we're, when we're talking about this specific models that are interacting with networking data, can you talk a little bit about the model you're using and source of training data? Yeah, excellent question. Uh, what model are we using?

We are get, uh, we get asked this question. Yep. All the time.

We are using state of the art off the shelf models. Okay. We are not training so Like oh seven B or something like that.

We, we are not training our own models. Mm-hmm. And, and that's a very intentional choice.

Mm-hmm. Because we want our customers to be able to plug forward AI into their own models that they can host. Okay.

Uh, so yeah, so it's, it's completely off the shelf. It's not, it's not fine tuned or, or trained with any specific data. Okay.

So the, so the model is really just being used for the natural language interface much less than for, uh, looking at the actual network. The understanding of what the network is doing comes from the forward AI part of it. And the mod, the AI is really sorry, comes from the forward networks part of it.

The AI is really just, uh, a user interface enhancement. It's Doing two things actually. Like that's, that's exactly one point, like, which is understanding this, this conversational pattern, natural language understanding and producing the result.

The other thing that the model is doing is also planning a sequence of tasks that it can do based on what capabilities are available from, from the four platform. So do you provide fine tuning of LLMs or, or, you know, if I wanted to go to LAMA three or LAMA four, whatever like that for my LLM in this case, how would I go about that? I mean, what's it, what's it look like?

Yeah, so the, the system is, it can plug it into a model that you host. It can, like right now what I'm showing you, uh, is the forward AI is connected to a model that we are running in AWS bedrock, but it is configurable. It can connect to, uh, a different model that you would be hosting.

But again, just to be very clear, uh, you can't just plug in any arbitrary model. Mm-hmm. There are, like, again, like not all models are the same.

Some work, well, some don't work well, uh, we can be at using, I'm confused now. So the, um, the model that's being used for the, the natural language Yeah. Is not the customer has to bring their own or that you provide it.

We have a built-in, we have, we host our own model that the SaaS platform can connect to, but the customer has an option to connect forward AI to their own model if they, if there's so desire. Okay. So in the one that you guys have, can you tell us what model you've Used?

Yeah. Happy to. Yeah, so we are using Tropics, uh, we are using two models.

In fact, we're using two of Anthropics models. There is a, there's a big model, the Sonet model, and then there's a small model haiku. So we are using a combination of those uhhuh to, to produce these results.

And so there's obviously some kind of fine tuning done so that it understands everything that your digital model can, um, duplicate, can be digital of that it brings into that system, into that new model. And that's what it's communicating with is with itself, with you with the user, but it's, it's using the data from the digital model and how you've trained it. So like that, that philanthropic, have you trained it down to We have not.

We are using it as it as it is, as Is. So are, are you, is there a way to put, are you, is there guardrails on that? How are you putting guardrails On?

The guardrails is completely on the agent side that, that it uses to figure out whether the question that is asked is in scope or not. It's like all of that, all of that logic is handled by the agent. Okay.

Not by the model. And Okay. So if you guys update the model, is that, are you planning on having to do that, you know, as the model changes in the future?

Yeah, absolutely. And then, and like I, we don't think this is the only model that works. Like there is nothing special about these specific models.

Like we, we, we, it'll, it should work with multiple other models and we will give options to our users to pick the model that works for them. And you supply the agent, or is that something that, uh, it comes out bedrock or cloud or whatever, or where, where's the logic that goes through and says, okay, we need to check this configuration, we need to look at a network. We need to go make some, some potential changes.

At least track the changes. I mean, how, where is that logic redesigned? That is the agent that is part of the forward platform That, that the demo that you saw.

So all of that That's developed the agent, yes. Yes, exactly. All of that lives within the forward platform.

Okay. Mm-hmm. And so the agent has guardrails on it then?

Yes. Okay. So the agent is checking, is this a task I know how to do and then it punts it upstairs if it needs to.

Exactly. But that's how you're controlling so that the generic model is not just making it up out of whole cloth. Yeah.