Cloudflare launches firewall for AI. This solution is intended to protect LLMs from abuse by organizations using Workers AI.

Cloudflare is introducing a so-called firewall for AI. Despite the name, this solution is not a true firewall, but rather a kind of new security layer designed to detect and block attacks and abuses of Large Language Models (LLMs) before they can cause damage.

The solution must protect AI applications that interpret complex data and human language. Based on research, Cloudflare finds that only 25 percent of executives have confidence in their company’s AI security measures. This solution is intended to help increase trust.

Security challenges for LLMs

Developing adequate security systems for LLMs is a complex challenge, according to Cloudflare. Models are vulnerable to misuse, attacks and manipulation, among other things because they can generate different outputs from the same input.

Cloudflare’s AI Firewall provides a solution by quickly detecting and automatically blocking new threats without the need for human intervention. Additionally, Cloudflare offers this protection free of charge to any customer running an LLM on Cloudflare’s Workers AI. The aim is to improve the security of AI applications and prevent problems such as prompt injection and data leaks. In short, the “firewall” checks prompts regardless of the underlying LLM and the tool blocks suspicious queries.

Useful, but not perfect

The solution seems quite useful. Generative AI models are indeed vulnerable to rapid engineering. Hackers have already proven many times that they can, for example, bypass ChatGPT’s built-in security to force the model to do or say things that should actually be protected. The more private and sensitive company data is available to AI models, the more important it is to prevent such misuse.

On the other hand, the reliability of LLMs and generative AI is more related to the training on the one hand and the quality of the data on the other. Without training based on qualitative data, you can never fully trust an AI model. Cloudflare does not address this fundamental challenge with the firewall.