With the global, regular use of generative AI nearly doubling over the last year, according to McKinsey, rapid adoption has created a new lucrative target for cybercriminals. While ‘off-the-shelf’ solutions have made up a great deal of this adoption, organizations seeing the power of fine-tuned, business-specific responses have directed a great deal of budget towards training their own AI models.
Continuous innovations, like agentic AI, mean adoption is only increasing. The considerable autonomy agentic AI possess allows it to can make decisions, plan actions and learn from its experiences within the specific context of a business, making it applicable across business functions.
However, amid the excitement around AI are considerable cybersecurity risks that all too often aren’t being considered. By adopting any new software solution, businesses are introducing a new attack vector for cybercriminals. The problem with in-house developed AI models is that they are essentially a repository for a company’s most valuable data, ranging from intellectual property, customer and employee data, and trade secrets, making it a highly attractive target.
This software runs off hardware likely housed in a data center, so business leaders need to ensure they are equipped with the right tools to have control over all aspects of their network to ensure sensitive company data is safe in the new attack vectors they are adopting.
CEO and co-founder of Goldilock.
The challenge to existing frameworks
The fact is that many businesses’ existing security setups aren’t currently fit for purpose. For years, IT departments have viewed cybersecurity as a compliance hurdle rather than a way to protect company data. This has led to an over-reliance on perimeter defenses and single sign-on solutions, which can create a false sense of security for organizations that believe compliance is equal to security.
Software solutions and more traditional approaches to data security, such as firewalls, still have a place in protecting a company’s data security, but a greater depth of defense is required to ensure operations run smoothly. AI, while powerful, is still a type of software running on hardware typically found in a data center. Data centers are complex and sensitive environments. Factors such as power requirements, cooling systems, and physical security make these facilities prime targets. Moreover, the nature of AI development and deployment requires frequent access and updates. This necessitates strict control over who can access these systems and when. Organisations need to ensure they have the right framework in place to ensure their AI models run correctly and are protected at all levels of operation.
Physical segmentation: establishing control and defense
Many will already have some of the components required in place. What the majority are missing is a first and last layer of defense which can be establish via physical network segmentation. Through a hardware-based approach, physical network segmentation enables users to segment all digital assets remotely, instantly and without using the internet. Through the press of a button, from anywhere in the world, organizations can use this technology to physically isolate their chosen segment from the overall network, disconnecting it from the internet. This technology acts as a guardian for AI, controlling access and ensuring its benefits can be reaped. For businesses using AI, it can offer the following benefits:
1. Improved security and reduced risk
In the context of protecting an AI model, this type of protection can act as a guardian, preventing a business’ own AI being poisoned, and preventing the use of AI for malicious purposes.
With no connection to the internet, physical network segmentation can be used to disconnect the model, preventing a cyber-attack or unwanted access. This will hide assets from view and enhance an organizations’ existing depth of defense. For AI models, network segmentation can be used to keep components offline until needed, massively reducing the window of time a hacker has to access the software.
Organizations may be hesitant to adopt this approach, assuming it would cause interruptions in operations. But this doesn’t have to be the case. The key is implementing a process that lays out clever and well-considered timing. A generative AI model doesn’t necessarily need to be connected to the internet 24/7 to perform well. A connection is required during a short window when users send a prompt. Once sent, the model can be disconnected and reconnected once the response has been generated and needs to be sent back. This short period of time is not nearly enough for a cybercriminal to clone the model and get their hands on sensitive company data. In terms of user experience, the time taken to connect and reconnect should be short enough that humans will not be aware of a delay.
2. Aiding regulatory compliance
Governments worldwide are adapting to the sensitivity of data. With AI models housing such an array of sensitive data, all eyes are on businesses to prove they are doing everything possible to prevent an attack or breach. With a lack of AI specific regulation, it’s hard to know where to start. Physical network segmentation can support overall compliance because there is no better effort than keeping sensitive data completely off the internet or physically separating it when attacked.
3. Effective incident response and recovery
In the case of a cyber-attack, reactive network segmentation can be used to impede attack propagation and isolate compromised assets and data quickly, effectively preventing further access for hackers. During the recovery process, leaders will have the ability to then rapidly reconnect previously isolated, known safe, segments after an attack making it possible to ensure AI models can be used as soon as possible and ensure the restoration of services.
Looking ahead
With more and more AI models trained in-house, cybercriminals will more than likely start to target these repositories of sensitive data. Once they have access to the AI, all sorts of havoc can be caused by the ability to clone the data, poison the model to generate harmful responses or lock it down with ransomware, causing significant company damage.
Organizations need to be able to confidently leverage the power of AI without compromising on security. By implementing a framework that allows the individual control of zones through network segmentation, business leaders will be able to not only mitigate threats, but also establish effective response and recovery processes while ensuring maximum performance business wide.
We’ve set up a comprehensive list of the best AI tools.
This article was produced as part of TechRadarPro’s Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro
