Assail Ships Its Own AI Model to Customers, Keeping Vulnerability Data Out of the Cloud

Watch  ·  Listen  ·  Read

Federal, military and large enterprise buyers want to test their own systems the way an attacker would, and many are uneasy about doing it in the cloud. Alissa Knight, CEO & Chief AI Officer at Assail, says her company answers that worry by shipping its own AI model to the customer, so the work stays on the customer's hardware.

Why are buyers wary of running offensive cyber operations in the cloud?

Cost and data exposure are the two concerns Knight hears. She says Assail sells to federal, military and large enterprise customers, not to small and midsize businesses, and that "the prevailing concern right now is prosecuting offensive cyber operations in the cloud." Token costs, the usage charges for cloud-hosted AI models, are a real worry for enterprises at the moment, she says.

The worry is not limited to government. Knight says insurance, financial services and fintech organizations she has talked with find it extremely cost prohibitive to prosecute offensive cyber in the cloud.

What does it mean to get the whole stack, model included?

Knight on what wrapper platforms do with customer vulnerabilities

Knight says Assail is the only adversarial exposure validation platform that gives the customer the entire application stack, including the AI model. She describes the category plainly: it is "basically fancy marketing speak for an autonomous hacking platform." In her telling, the model carries 26 years of her own offensive experience with APIs and web applications, written into training files and turned into small tuning layers on a base model.

Her contrast is with other products. She says many are wrappers around frontier models, and that those wrappers send customers' vulnerabilities to the cloud. That is her claim about competing products, and we have not verified it.

How small can a model be and still run an attack platform?

Model size as Knight describes it

Knight says the model is about 60 gigabytes with 31 billion parameters, and that it runs on anything with enough video memory. She names the RTX 6000 Pro as one option and says customers do not need to buy H100s, B100s or B300s for inference, the step where the model produces its answers. She adds that Assail has just ported the model to ARM, so it runs natively on the Grace Blackwell chip as well as on x86. These figures are hers and are unverified.

Knight says the same approach covers targets well beyond web applications. She describes training material on industrial control systems, drones and embedded operating systems, and gives VxWorks as an example of an embedded system that appears in weapons platforms.

Why does a model's rejection rate affect what a customer is told?

Knight says frontier models have a high rejection rate when asked to attack something, because guardrails are built into them. In her account, a platform built on one can report zero vulnerabilities when the real reason is requests being rejected in the backend.

She makes that point without naming a competitor, and it is her claim. Her conclusion for buyers is direct: "the days of just trust the vulnerability scanner are gone." She says Assail's reports include the exploit evidence, a video of the attack and a patch the customer can hand to developers, so the operator can see a finding is real.

A clean report may not mean a clean application, in Knight's account

What has running disconnected opened up?

Running without an internet connection has let Assail sell into the federal and military space, Knight says, supporting warfighters in contested areas who have no connection. Customers also get what she calls AI sovereignty and data sovereignty. Assail has spun its federal group off as Assail Federal.

Knight also says that the company went from zero customers, zero revenue and no funding to its current position in nine months, and that all of its revenue so far has come from people who attended last year's ISC2 Security Congress, where she gave the keynote. Those are her figures.

For a buyer weighing any platform like this, her account suggests three questions. Where does the model run, and where do my findings go? What does the report say when an attack request is refused? What hardware does it need? The full conversation with Knight and Sean Martin is available to watch and listen to, and Knight invites readers to connect with her on LinkedIn.

This is a Brand Highlight. A Brand Highlight is a ~5 minute introductory conversation designed to put a spotlight on the guest and their company. Learn more: https://www.studioc60.com/creation#highlight

Questions answered in this article

Why are federal and enterprise buyers wary of cloud-based offensive cyber operations?

Alissa Knight, CEO & Chief AI Officer at Assail, says the prevailing concern is prosecuting offensive cyber operations in the cloud, and that token costs are a real worry for enterprises. Read more: Why are buyers wary of running offensive cyber operations in the cloud?

Which industries does Alissa Knight say find cloud offensive cyber too costly?

Knight says insurance, financial services and fintech organizations find it extremely cost prohibitive to prosecute offensive cyber in the cloud. Read more: Why are buyers wary of running offensive cyber operations in the cloud?

What does Assail mean by giving customers the whole stack, model included?

Knight says Assail is the only adversarial exposure validation platform giving the customer the entire application stack, including the AI model. She says many other products are wrappers to frontier models that send customers' vulnerabilities to the cloud. Read more: What does it mean to get the whole stack, model included?

How big is Assail's AI model and what hardware does it run on?

Knight says the model is about 60 gigabytes with 31 billion parameters and runs on hardware such as the RTX 6000 Pro, without H100s, B100s or B300s. She says it has just been ported to ARM for the Grace Blackwell chip. These figures are hers and unverified. Read more: How small can a model be and still run an attack platform?

Why does a frontier model's rejection rate matter for vulnerability reports?

Knight says frontier models have a high rejection rate for attack requests, so a platform reporting zero vulnerabilities may be hiding backend rejections and giving customers a false sense of security. It is her claim, and she names no competitor. Read more: Why does a model's rejection rate affect what a customer is told?

What has running without an internet connection opened up for Assail?

Knight says it has let Assail sell into the federal and military space to support warfighters in contested areas, and gives customers AI sovereignty and data sovereignty. Assail has spun its federal group off as Assail Federal. Read more: What has running disconnected opened up?

For all of the content in this series, have a look at the main series post: Sean Martin is Live from ISC2 Security Congress 2026.

Read the full transcript · 1,639 words

Alissa Knight - Assail Interview Edited MC

Sean Martin: let's do huggers. Ah,

Alissa Knight: alright.

Sean Martin: So good to see you.

Alissa Knight: Good to see you again too.

Sean Martin: And to see all of this.

Alissa Knight: I know. It's been, it's

Sean Martin: super impressive.

Alissa Knight: It's crazy. I've, I've never, so, as you know, I've started and sold multiple cybersecurity startups. I've never built a company so fast like this all happened in nine months. I was the keynote speaker last year in Nashville at ISC2 Security Congress. We had zero customer, zero revenue, uh, no funding. We had 300 signups when I was on the keynote stage. 100% of our revenue so far has been from people that attended last year's conference, so ISC2 has been a very successful venue for us.

Sean Martin: That's

Alissa Knight: good. Been for us. So

Sean Martin: it's a great testament.

Alissa Knight: Yeah.

Sean Martin: Well, Alissa, it's great to see you.

Alissa Knight: Thank you.

Sean Martin: And, uh. So we're gonna talk a little bit about some of the stuff you're seeing, some of the trends you're seeing.

Alissa Knight: Yeah.

Sean Martin: That led you to some announcements as well.

Alissa Knight: Sure. Yeah.

Sean Martin: So give us an overview of what, uh, what you've been up to.

Alissa Knight: Yeah. So we've, in meeting with federal, military, large enterprise, we don't sell to SMBs. The prevailing concern right now is prosecuting offensive cyber operations in the cloud. Tokens is a real concern for enterprises at the moment. We've talked to different types of organizations, different industries, not just government and military, but uh, insurance, financial services, FinTech. There it is extremely cost prohibitive to prosecute offensive cyber in the cloud. And so as the only AEV platform that provides the customer the entire application stack, including the AI model, we are cornering the market right now. Uh, a lot of the other competing solutions out there are simply just wrappers to frontier models. So the wrappers to Claude, the wrappers to ChatGPT, they are sending your vulnerabilities to the cloud. I don't understand how that's ever been a thing, how that's even. Okay. And so with our platform customers get that AI sovereignty, they get that data sovereignty, and it also allows us to run in disconnected environments. So we don't have a connection to the internet. You can use our platform. So this has really, you run

Sean Martin: on the field.

Alissa Knight: Yeah. It's, it's allowed us to really, uh, sell into the federal military space to support the war fighter who is in contested areas. They don't have a connection to the internet. And because we're a small language model, we're only about 60 gigabytes in size with 31 billion parameters. You can really run us on anything that's got enough B ram. So you can run us on RTX 6000 Pro, you don't need to buy these expensive H100s, B100s. Uh, B300s for inference. As a matter of fact, we just ported our model to arm architecture, so we actually run native, the new Grace Blackwell chip. So we run on X 86 and arm as well.

Sean Martin: Very cool. So I would imagine a lot of people kind of understand the importance of this,

Alissa Knight: right?

Sean Martin: And maybe give us a scenario or two where it's critical to have the, the model inside and on-prem

Alissa Knight: Yeah.

Sean Martin: In in the area where you want it to be and not off somewhere else.

Alissa Knight: Yeah, sure. So just. For your audience's edification. We're an adversarial exposure validation platform, so that's basically fancy marketing speak for an autonomous hacking platform. So I basically put myself out of a job. Right. As a pen tester.

Sean Martin: It is you?

Alissa Knight: Yeah. This is pretty much, yeah. I mean, all my 26 years of offensive experience in hacking APIs, web apps, we baked into JSONL files, which became QLoRA adapters for the model, and so. Uh, the reason why ha shipping our own AI model to the customer is, is important in several, several ways. Number one, you have a high rejection rate with those frontier models. So when a competing platform, uh, fill in the blank on who that is, uh, is reporting zero vulnerabilities. Sorry, I didn't find any vulnerabilities in the application. A lot of the times it's because of a rejection rate that's happening in the backend that they're not telling the customer. So because they're just the, the scaffolding or harness around, uh, that other model that they're borrowing, there's a high rejection rate with that. 'cause it's really difficult to go to something like Claude or ChatGPT, which has those guardrails baked in to say, I want you to attack my API swear this is legitimate. I own it. It's there's, you're still gonna get rejections,

Sean Martin: right?

Alissa Knight: So it's going to affect your findings and you're gonna get this false sense of security that. Hey, I'm not vulnerable. There's no vulnerabilities. Even AI couldn't hack me when in fact it was just because of a high rejection rate with a lot of those platforms.

Sean Martin: So interesting. So you mentioned [00:05:00] kind of taking your brain, right? Putting in JSONL files there, you said.

Alissa Knight: So yeah, so I mean, so JSONL files are effectively the recipe, if you will, for creating your QLoRA adapters, which is basically weight. So you're basically taking a model and you're fine tuning it. They're think of them like dial. So you're taking a base model and you're redialing its focus, I would say, or expertise. And, uh, for us it's offensive security. So we've cr effectively created training files, our own training files on this tacit knowledge or corpus of data on how to hack SCADA systems, industrial control systems, um, drones embedded operating systems like Wind River, uh, that sort of thing that our connected world relies upon. So like, for example, the. VxWorks is the operating system for the Apache attack helicopter. It's in submarines, it's in the, you know, joint strike fighters. It's in all of these weapons systems. Um, and it's an operating system. It's an embedded OS, and so our platform can hack those targets, so it gives the warfighter federal agencies who want to attack their own assets and also enterprise that wants to attack their own assets before the bad guy does. It's identifying the vulnerabilities. Exploiting them and, uh, being informed about those vulnerabilities before the bad guy finds 'em first.

Sean Martin: And let's wrap with what comes out the other side. 'cause it's great to have findings.

Alissa Knight: Yeah. You know,

Sean Martin: maybe

Alissa Knight: not great to have findings. Yeah. What do we do about it? Right. So the, the interesting thing is, so first and foremost we have post quantum encryption. So we shard our encryption keys up, making it impossible for a quantum computer to crack our encryption. So everyone, but. The outcome is, uh, an interactive report. It's explained in plain English, the findings. We provide all of the exploit response, so it's, you know, the days of just trust the vulnerability scanner are gone. We actually give the findings from the vulnerability we exploited, and so it really gives that user or operator the confidence that this is a real vulnerability, real findings. She's actually pilfering or capable of pilfering, um, data from the target to prove that she was there. She also records a screen, so when she's attacking the target, you actually get a video of the attack. So, and then finally patches. So she will actually patch the vulnerability. She'll give the user the patch, the, the user can then provide that for their developers to actually patch the system. And the new Charlemagne release that we're releasing on December 31st. We're getting more into actually memory resident patches. So instead of modifying the source code or trying to patch the system, the application itself, especially when Ed doesn't have access to the code, you can create a patch that runs strictly in memory. Very

Sean Martin: price mitigation. Yeah.

Alissa Knight: Yeah. It's very cool.

Sean Martin: Super cool. Well, I'm so excited for you.

Alissa Knight: Thank you, [00:08:00] Sean.

Sean Martin: Assail is a company.

Alissa Knight: Yes.

Sean Martin: And this is your name?

Alissa Knight: Yeah. Yeah. You're

Sean Martin: amazing

Alissa Knight: that that was an accident, by the way. But yeah, so. Someone walked up to me and said, that's really cool what you did with the name. I was like, what are you talking about? You know it's an anagram for your name Alyssa. Right. I was like, I, no, I did not know that. And so you just said it. Uh, so yeah, it's actually from Sun Tzu's Art of War, the book. Exactly. Uh, and you know, it's something along the lines of, I'm gonna master this quote, but somewhere along the lines of, we don't anticipate, it's not that we're anticipating the enemy not coming, but that we've made our position unassailable. So it's unassailable. Yeah. To attack. I

Sean Martin: love it. I'll tell folks how they can reach you.

Alissa Knight: Yeah. So, uh, come visit our website. It is assail.ai.com. We spun our federal group off into Assail Federal, so that's assailfederal.com. And, uh, you can connect with me on the socials, like LinkedIn and X. So

Sean Martin: do it. She's amazing. Thanks everybody.

Alissa Knight: Appreciate it. Thank you. Awesome. Thanks Sean. Appreciate it's good seeing you again.

Sean Martin: So