Centralized protection against AI Usage Risks with Network-based AI firewall

By: Brett Helm

Intro 

Generative AI solutions, such as ChatGPT, Google Gemini/Bard, and Microsoft Bing AI/copilot, have exploded in popularity. The adoption of ChatGPT has scaled so quickly that users frequently find that the system is at capacity, and they must wait for access.

AI solutions can improve efficiency, reduce errors, and improve an organization's operations. There is, however, a dark side as these tools create a risk to the enterprise. There is a risk that enterprise employees will use AI vendors without authorization and without enterprise agreements in place to protect confidential data. Guardrails are needed to protect against the obvious risk of confidential data being leaked. AI solutions pose additional risks to the enterprise including inbound malware, prompt injections, and copyright violations.

Last year, Samsung banned the use of AI tools from company devices and personal devices connected to company networks, after a data breach. Data was leaked to ChatGPT in three separate incidents. In one case, an employee asked ChatGPT to generate notes from a recorded meeting. Another employee asked the chatbot to check source code from a sensitive database for errors. A third employee input code into the chatbot and requested ChatGPT to provide code optimizations.

 

Generative AI and Large Language Models

AI solutions are being adopted by companies, both large and small. These tools can be used to automate a wide range of repetitive tasks and increase efficiency. Use cases for AI include:

• Software developers are using these tools for code generation, documentation, and debugging.
  
• Marketing teams are using them to generate presentations and social media content.
• Managers are using them for data analytics and report generation.
• Manufacturing companies are using them for predictive maintenance.
• Customer support organizations are using them to route emails to the right groups, and even to provide automated responses.

ChatGPT and its analogs are AI based chatbots, built on top of a Large Language Model (LLM). Large Language Models are a type of neural network designed to process and generate data in sequences. In other works, they process text-based input and generate responses based on the input they receive.

LLMs can write poetry, summarize meeting notes, generate a PowerPoint presentation from a text description, or analyze software programs to discover errors.

One of the main reasons that AI tools such as ChatGPT, Google Bard, etc. are so powerful is because they were trained on a massive data set.  These systems scrape information from the Internet and use that data to train the LLM. 

 

The diversity and depth of information available on the Internet is larger than most people realize. LLMs have seen billions of samples of writing on almost every conceivable topic. Hence, we can think of a LLM as being trained to produce text that could reasonably be expected to appear on the Internet. This covers everything from song lyrics to legal precedence, and from software code to textbooks and homework assignments. This breadth and depth of input material enables the LLMs to generate sophisticated and complete responses on almost any topic.

Data breaches

Given that AI tools utilize gathered information to provide the user with an accurate product, the user must never forget that the AI system is always mining information. Even proprietary information, once in the system, will be seen and used by others. 

The ChatGPT data policy states that data input into the chatbot will be used to train its models, unless users explicitly opt out.

This is a serious concern for both the company's confidential information and for personally identifiable information (PII) managed by the company. Should this information be leaked via an AI tool, a company could find themselves in violation of legislation including GDPR and the California Consumer Privacy Act (CCPA),

Unless one exercises care, it is very easy to unknowingly leak data with ChatGPT or one of the other AI tools. Employees are experimenting with the capabilities of the tool and can easily cut and paste data into the web interface without realizing the risks.

Copying and pasting data into an AI web interface is not the only avenue by which organizations can accidentally leak data. AI tools also provide API access, allowing companies to build applications to automate workflows that leverage the power of these AI tools.  Automation is a powerful tool to improve worker productivity, but it removes human control. Any data input into the automation process could be leaked through the APIs and added to the training data for the LLMs, and there is no one watching the data to ensure confidential information is not fed to the chatbot.

Enterprise Agreements are not a Silver Bullet

ChatGPT offers an enterprise license that, among other things, ensures protection of a company's data. With an enterprise license from ChatGPT, companies retain ownership and control of their business data. The system won't use business data or conversations for training of their LLM. 

Enterprise agreements offer a solution to concerns over data breaches from using AI solutions, but only to the extent that they are used. A company must ensure that all access to AI solutions uses the company's enterprise account. If an individual uses a personal account, there is still a high risk of data leakage.

Types of AI Firewalls

Despite being a relatively new category in the market, there are already multiple types of AI firewalls.

• AI Prompt Firewall: A browser-based plugin that manages and controls an individual's access to AI tools. These tools only monitor access to AI tools through an AI prompt accessed by a browser. These tools don't provide any monitoring or control over API-based access to AI tools and are specific to a single browser. A plug-in must be installed on each device within the enterprise, and a new version is required for every browser utilized within an Enterprise. There are also challenges with federation, centralized policy management, and centralized control.
  
• AI Powered Firewall: Tools using AI/ML to create a better Next Generation Firewall (NGFW). This is really a different solution altogether, as it is focused on using AI to create a better firewall. It is not focused solely on enforcing appropriate use of AI tools.
• Network-based AI Firewall: Much like traditional network firewalls, these tools sit between the corporate network and the external internet and enforce corporate communication policies. A network-based AI firewall can be deployed behind the existing firewall to provide additional protection specific to the use of AI tools.

A network-based AI firewall is the only solution that provides full protection against the misuse of AI tools across the entire enterprise, including prompt-based and API-based access to AI solutions.

Network-based AI Firewall

A network-based AI firewall has visibility into all of an enterprise's network traffic being sent to AI solution providers. With this broad purview, the firewall supports the following characteristics:

• Discovery of AI usage
• Reporting of AI usage
• Lane classification
• Guardrails
• Quarantine

 

An AI firewall enforces company policies regarding usage of AI tools.

 

Discovery of AI usage

Network-based AI firewalls have visibility into all AI requests. This allows centralized discovery across the entire enterprise, regardless of the AI vendor, or method of access. Browser-based plug-ins don't provide centralized data collection and only have visibility into prompt-based access to AI tools.

Reporting of AI usage

AI usage collected by network-based AI firewalls enables analytics and reporting on an organization's AI usage. This can be used for:

• auditing AI usage vs. billing from AI vendors
• monitoring for inappropriate AI usage or AI usage that does not follow company policies
• developing metrics on the adoption of AI across company departments and functions
• measuring department performance improvements based on AI usage

Enterprise-wide reporting can only be achieved using a centralized-federated AI discovery solution.

Lane classification

Many enterprises will adopt a specific AI solution based upon a specific use case, or "lane". For example, an enterprise may choose one AI company for the "software development assistant" use case. A separate AI company may be approved for content generation, image generation, or voice generation.

Lanes could be broadly designed, or narrowly designed, depending on the needs of the enterprise.  For example, a large bank may contract with Microsoft for a software development assistant to assist developers using a scripting language. The marketing department may then contract with OpenAI for content creation. Alternatively, a technology company may employ multiple lanes for software development groups using different programming languages.

Guardrails

Guardrails enforce compliance AI lanes, ensuring compliance with corporate policies on AI usage. Security guardrails could be used to protect against the loss of privacy data, and non-approved AI access. Combined with quarantining, they also provide security capabilities.

Quarantining

Quarantining capability allows organizations to implement fine-grained protection from unsafe scenarios such as data leakage or malicious code responses from an AI company.

The Solution

Companies must set policies on the use of AI solutions. These policies must balance the benefits of using this new technology with the risk. 

Once a company policy has been set, employees must be trained on the policy and the acceptable use of AI solutions. Ideally, the company will invest in an enterprise license to enable broad usage of AI while managing its risks.

Finally, companies must enforce the policies they have created. A network-based AI firewall can be used with a customer's NGFW to make fine-grained AI policy decisions. An AI firewall could be installed inline behind a traditional firewall to block requests that don't comply with company policies. The solution can also be installed alongside a traditional firewall or network router via a span port or transparent traffic ingestion. With this configuration, the solution acts as a visibility and discovery tool, providing detailed reporting on usage of AI tools.

Utilization of a network-based AI firewall provides several benefits over browser-based plugins:

• The solution is easy to install and centrally managed
• The solution provides a single point data collection for AI usage audit and privacy compliance enforcement
• No changes are required to user endpoints
• The solution manages all access to AI tools, including access from any browser or via APIs for cloud-based connections or on prem solutions.
• Protection for non-browser-based applications
• Supports both cloud and legacy on prem environments.
  

A network-based AI firewall will ensure that all access to AI solutions utilize the company's enterprise account. Attempts to access AI solutions that don't use the enterprise account are either blocked or redirected to the enterprise account.

Conclusion

AI Solutions are transforming how businesses operate. ChatGPT alone has been adopted by over 80% of Fortune 500 companies. Companies are still learning how to use AI and adoption will continue to grow with time. New use cases will be discovered, and companies will refine and optimize their use of this technology. But the use of AI Solutions is not without risk.

By default, information provided to these solutions is added to their LLM, allowing them to continue to learn and adapt. Unless care is taken, a company's confidential and private data will be added to the LLM, making it available to anyone who queries the system. 

Companies need to develop policies on the use of LLM based systems to ensure their private data remains secure.  They must evaluate the risk, set policies, educate employees, and ensure the proper controls are in place to enforce the policies they set. An AI firewall provides the protection required to enforce these policies.

The use of an enterprise licensed AI solution is not a silver bullet. An enterprise license will ensure that confidential information is not added to an LLM, but only if it is consistently used.  Companies must implement controls to ensure employees use only the enterprise accounts. An AI firewall allows enterprises to take matters into their own hands and control risks associated with AI usage.

##

ABOUT THE AUTHOR

 

Brett Helm is the Co-Founder and CEO of Glasswing, a provider of the industry's first automated AI discovery and AI firewall platform. This solution is available now. Previously, Brett held CEO roles at DB Networks, Coradiant, and iPivot, Inc., as well as senior management roles at Intel.