G&L CEO Alexander Leschinsky Talks Streaming Security Attack Vectors and Best Practices in the Age of LLMs
In this exclusive interview, G&L Systemhaus CEO Alexander Leschinsky explores the evolving landscape of streaming and content security, highlighting the increasing sophistication of AI-powered attacks, and the importance of integrating security into all layers of media workflows, while outlining a modern security stack that includes DNS security, web application firewalls, DDoS protection, securing high-risk APIs, access control, and testing to identify vulnerabilities before attackers do.
AI-Powered Attacks
Nathans-Kelly kicks off the discussion with a question about the evolution of streaming security and what types of high-risk threats streamers face even when they see themselves as relatively secure.
“The most important factor is AI right now, because we all know how it helps us to work and smooths our workflows, but it helps the good people exactly as it does the bad people,” Leschinsky replies. “If you are an attacker, you can use AI to easily design very complex and very sophisticated attack vectors into content owners' repositories. The move to AI-powered attacks is a total change from just being an opportunistic attacker that just uses the knowledge that they naturally have to somebody who's doing this more systematically with the help of an expert. That means that the AI-powered attacks are much more sophisticated, can be replicated on a much larger scale.”
As a result, he goes on to explain, “It's very cheap now to attack on a large scale and to be very sophisticated. That means that AI has a huge effect on the elaborative nature of these new kinds of attacks that we see, and that has real consequences for customers in terms of the rights that they acquire, monetization, and brand trust. So protecting the whole value chain against AI-powered attacks is what we see drives the most interest right now.”
Leschinsky also points out that the guardrails that AI vendors put in their products don’t necessarily act as a deterrent to attackers who leverage them; “there are many ways to counteract and walk around that.” What’s more, he says, “There are so many open-source large-language models where you can work without any of these boundaries. So it's an illusion to think that AI is not helping the bad guys.”
Minding the Security Gaps
Turning the discussion toward areas where security is most likely to break down, Nathans-Kelly asks, "What are the most common security gaps that you still see even in platforms that are well-run?"
Again, Leschinsky points to AI: “The first one and the newest is attacks that attack your usage of LLMs yourself,” he says. “So if you are using AI and you have a chat engine—either a public one or one of your own that is running on- premise in your company—that is good and helps you a lot. But as soon as these agents have access to the public internet, they are at the risk of getting prompt injection, getting false information that can be fed into the answer that they generate for the prompt. That’s something that we can help shield you from.”
APIs are also a critical area where security gaps occur. “Everybody is using APIs. Every company is API-first by now, but the problem is that these APIs spawn up so quickly and are often so unstructured and un-unified that it's even difficult to have an inventory. We ask customers, ‘Hey, how many APIs do you have?’ They say, ‘We have five, six, seven APIs.’ And then if you do a detection, you see they have hundreds of APIs. And so the inventory of APIs is very important because with the APIs, you can do a lot of harm and you can get access to a lot of content. So API access is a common security gap.”
The third crucial gap, he says, is “bot traffic that tries to either grab content that is not legitimately theirs or bots that try to get information that they can use otherwise or that they can resell or use on their own portals. Bot traffic is a huge issue, especially because the bots get more and more sophisticated,” which makes it difficult to determine what is a legitimate user and what is a bot.
Difficult as it is to make that distinction, he continues, “it makes a huge difference if you can detect them.”
The last gap he cites is “attack chains, where you combine different vectors. Again, that is easier with AI today. So you can take a lot of different approaches to attack content and to get access to APIs or to secure content, combine them, and then it's even more difficult to shield them off.”
Beyond the Firewall
“Even with the omnipresence of bots and the more systematic attacks that we’re seeing,” Nathans-Kelly says, "I imagine there are still content owners who assume they’re secure enough. They say, ‘We already have a firewall, DDoS protection, and good people working on it on our end.’ What do you tell them?”
“Of course it’s important to have a firewall,” Leschinsky concedes, “and this is the basic 101 of IT security, but it's based in traditional parameter thinking. So you’re thinking you're in your castle and you have your wall and nobody comes inside, but the problem is that modern attacks don't think about these parameters as in the past. Attacks usually disguise themselves as something that is a valid request,” he explains. “They walk through the gates that you provide and they do their harm pretending they were usual requests.”
The “pyramid of thinking” around security, he contends, has fundamentally changed due to these new attack vectors. “The firewall is not enough. You have to have a lot more sensors around to look into behaviour and into what seems to be legitimate-access traffic. [You need to] look at the details and detect the patterns that tell you if this is a bot or this is an attacker.”
Four Pillars of Platform Resilience
“Let’s talk a little bit about what a modern security stack would actually look like in streaming,” Nathans-Kelly asks. “Without simply rattling off a vendor checklist, what can you say about how that stack would look?”
“It all usually starts with DNS,” Leschinsky replies. “DNS is boring. It’s something that feels like the 1970s, but it’s the basis of everything we do. Every request we send starts with a DNS request. DNS availability is extremely important, and DNS issues have been the root cause of a couple of large outages that we had with hyperscalers and with platforms in the last months. So DNS is totally important. So secure your DNS set up in a way that is very professional.”
The next step, he continues, is “to add web application firewalls and DDoS protection. This is also very important. Make sure that whatever interface you have to the public, you secure it with a CDN that focuses on security and both for huge-demand denial-of-service attacks and for web application attacks that try to do database injection or steal secrets.”
In summary, he calls “web application firewall, DDoS, and DNS security… the first pillar of platform resilience” in a modern security stack.
Next, he warns content owners, “make sure that you have the kind of API integrity so that you always know what are your APIs that you and your teams and your distributed teams and your subcontractors that you work with, that they publish and open up to end customers or to partners. Do the discovery of what are your APIs and then monitor and protect these endpoints continuously as if they were your website or your crown jewels. So, API endpoint, discovery, monitoring, and securing is the second pillar that I would say you need in a modern security stack.”
The third pillar, he says, is access control. This includes “your user IDs, token and policy-based delivery. Access control and token authentication are very important beyond the question of if you have to use DRM or not. Many people use DRM and have to because the content owners demand that. Other customers don't need that because they shy away from the complexity and they want to provide their users with a premium user experience, and it's difficult to do that with DRM. DRM always adds an extra layer that you have to fulfill. Independent of that, access control is very important.”
Finally, he points to bot-scraping defense as the final pillar, “where you make sure that you can immediately identify if a request comes from a real user, a user that you want, that you want to send advertisers to, that you want to send your content to, who has paid for your content, or if it's just a bot that tries to steal content or credentials or something else. So bot scraping defense is the fourth pillar I think you need for a modern media security stack.”
What G&L Offers
Rounding to his last topic, Nathans-Kelly says, “We’ve talked a lot about where the threats come from and what the new threats in particular are that we face in our industry, as well as some best practices for dealing with them. But when security implementations do fail, let's talk about what G&L does specifically that's different from others in terms of integrating security into streaming workflows.”
"The way we usually start,” Leschinsky replies, “is by saying, 'Security is not a standalone silo that you can just add.' You don't say, 'I have media processing, I have media delivery, and then I have security.' It's not a separate entity. It's something that has to be deeply integrated in what you build and what you do and on all layers, both from a business logic perspective, and from a traffic delivery and operations perspective. For us, the first step is to make an inventory and see what are the different aspects in terms of business, traffic delivery, media processing, and operations, and how do they play together with this specific customer? And then we make a security landscape and see how these [elements] work together and which are the ones that we can secure in the easiest way with the least effort for the customer.”
Nathans-Kelly asks, “So, if a content owner comes to you and says, ‘We need to improve security this year,’ what do you tell them about how to achieve that and how quickly they’ll be able to achieve it?”
“So, I think the first step that is usually necessary for all the rest and that is easy to implement is measurable visibility,” he says. “Usually, that means that we increase the number of metrics and traces and logs that they gather so that we get real deep security-related and media delivery-related observability. If we want to increase and improve something, we have to measure it. So we usually start with just ramping up the lock levels and then looking at everything we see. And we have very good mechanisms in place where we can consume huge amounts of lock information, huge amounts of data, and pinpoint that and provide not only a huge amount of data, but also see what's important and where are the attack vectors that are hidden in the huge amount of data.”
The next step, he continues, “is to see which of the traffic is completely automated and which is more manual. And then we can discuss with the customer where their focus is, what we should support. So that first differentiation is usually helpful. And at the same time, we usually detect a lot of unmonitored APIs with these customers, but then we easily see which ones are the ones are most important or are most prone to attacks, and then we focus on a handful to start with [instead of trying to] cover everything at the same time. That doesn't make sense. We just pick out a couple of high-risk APIs and focus on them.”
A third approach he recommends is “looking at exposure surfaces” such as “admin access and identity. That’s also something where we can work like a pentester in a way, where we try to attack the system ourselves with the AI tools that an attacker would use, and then make sure that we find these holes in their systems before the attackers actually do this. So within our role as a systems integrator,” he concludes, “we can act as a security pentester as well.”
Related Articles
Interview with Alexander Leschinsky, CEO and Co-Founder, G&L Systemhaus
28 Aug 2025
Alexander Leschinsky talks content provenance, authenticity, C2PA and and DRM
07 Aug 2024
Companies and Suppliers Mentioned