Cloudflare Inc., a global cloud connectivity provider company, today announced the development of Firewall for AI to provide companies a layer of protection for artificial intelligence large language models, with an aim to identify potential attacks before they can tamper with critical functionality or access sensitive data.

The company also announced a suite of new defensive cybersecurity tools that use AI to fight emerging AI threats. Those include detecting anomalies in user behavior, scanning email to flag suspicious messages and mitigating threats to the organization.

As more companies begin to embrace LLMs and AI models as a central part of digital transformation, the enterprise must face the security risks that come with it. According to a recent Deloitte study, only one in four C-suite-level executives are confident that their organizations are prepared to face the risks that AI poses.

“We’ve entered into the AI arms race, and it’s critical that today’s AI-powered apps — many of which fuel our healthcare, our banking systems, and our grid — are built on secure models,” said Matthew Prince, co-founder and chief executive of Cloudflare. “This protection should be available to all because a secure internet is better for everyone.”

The new Firewall for AI will provide security teams the capability to rapidly detect new threats, the company said, as it could be deployed in front of any LLMs running on Cloudflare’s existing Workers AI offering. Workers AI allows developers to deploy AI models at the edge at scale on Cloudflare’s global network, which in turn puts any LLM running as close to enterprise customers as possible allowing extremely low latency responses.

By putting the firewall in front of the LLM, it can scan prompts submitted by users to identify attempts to exploit the model and extract data, the company said. As a result, it can automatically block threats as they happen, with no need for human intervention. Any customer running an LLM on Cloudflare’s Workers can take advantage of the Firewall and be safeguarded by this new capability for free to defend against growing concerns such as prompt injection and other attack vectors.

Prompt injection attacks aim to get an LLM to reveal sensitive information by carefully crafting a question that hijacks the way the model works and gets the AI to say exactly what the attacker wants. Using prompt injection, an attacker can override previous instructions given to the LLM by giving it new commands, which could then allow the attacker to get the LLM to reveal sensitive information or access critical functionality.

Cloudflare’s new protection uses AI to fight AI

With Defensive AI, Cloudflare said, it’s providing a personalized approach to securing enterprise networks against new risks coming from emerging technologies, such as AI-enhanced attacks, by using AI to detect emails, malicious code and anomalous traffic patterns.

According to the company, it has expanded its offerings to train AI models on customer-specific traffic patterns to tailor a defense strategy based on the baseline behavior of the company’s network and environment.

“By understanding ‘normal baselines’ in a customer’s environment, and mitigating the threats that will move the needle towards increased resilience, Defensive AI is the crucial edge defenders need to stay ahead of today’s adversaries,” said Prince.

With the rise of AI such as OpenAI’s ChatGPT, attackers have gotten a lot more sophisticated in phishing scams, which are a type of email or message that attempts to trick a user into giving up sensitive information. In the past, these scams were not always very convincing because the written messages were often riddled with glaring errors, such as grammar mistakes or poorly designed ones. But, now with the assistance of AI, attackers can tailor their emails to their targets to make them even more likely to be believed and get them to give up passwords or sensitive information.

With Cloudflare’s Defensive AI, email threats can be detected more quickly and thus scam messages identified before employees fall for them, the company said.

Under the same defensive approach, Cloudflare is developing an application programming interface anomaly detection model that will prevent attacks aimed at infiltrating networks, attacking applications and stealing data. The aim is to generate models of normal behavior within the network and then watch for anomalies in traffic, as attacks deviate from the proper behavior for applications, it will act as a guardrail against malicious attacks.

“We have entered into an era where fighting AI with AI is non-negotiable,” said Prince. “A personalized approach to protect data and defend against complex threats unique to an organization’s attack surface – at speed, and scale – is paramount.”

Image: Designpics/Adobe Stock