The Firewall for AI will function like a traditional WAF, scanning every API request with an LLM prompt for possible attack patterns and signatures. It can be deployed in front of models hosted on the Cloudflare Workers AI platform or any other third-party infrastructure. The firewall will also prevent volumetric attacks, identify sensitive information, prevent model abuses, and validate prompts and responses. The prompt validation feature is currently under development with a beta version set to be released in the coming months.
Key takeaways:
- Cloudflare is developing Firewall for AI, a protection layer for Large Language Models (LLMs) to identify and prevent abuses before they reach the models.
- Firewall for AI is an advanced Web Application Firewall (WAF) tailored for applications using LLMs. It includes tools to detect vulnerabilities and provide visibility to model owners.
- The Firewall for AI will run detections to identify prompt injection attempts and other abuses, ensuring the topic stays within the boundaries defined by the model owner.
- Enterprise customers on the Application Security Advanced offering can start using Advanced Rate Limiting and Sensitive Data Detection. A beta version of Firewall for AI’s prompt validation feature will be released in the coming months to all Workers AI users.