A firewall for AI agents

I had this idea the other day and decided to write about it to validate that people actually want the product before I start building. The idea is still very barebones, but I'd love some feedback as I get started. If you want to follow along, subscribe to get updates.

Problem

The rise of AI agents is introducing many unforeseen cybersecurity threats, especially for small business and enterprise customers who are simply trying to leverage this technology to streamline business processes. New and creative prompt injection attacks are popping up like wildfire, and most companies are eager to incorporate this new technology despite lacking the expertise to effectively protect their data and systems from these new exploits. Traditional cybersecurity experts without an AI background are not always equipped to independently handle these threats, and many companies are essentially opening up a backdoor into their organization without even realizing it. For example, here's a great video about a complex prompt injection attack that uses support tickets and the Supabase MCP server to leak sensitive user data and modify the database (this was actually the inspiration to tackle this problem).

These agentic AI tools are great, but they introduce wild amounts of security overhead for inexperienced organizations to deal with on their own. Model providers like OpenAI, Anthropic, and Google can't adequately protect against prompt injection attacks for specific business use cases because implementing this broadly across generalized models could lead to false positive flags that blowback on everyday consumers, so the responsibility for security is fully on these organizations. But it shouldn't be this difficult, and that's where I come in!

Solution

A plug-and-play standardized system for securing agentic AI systems - A firewall for AI agents

A middleware layer between the initial API request and the model provider adds extra security without extra complexity. This simple approach allows customers to easily tackle AI security by offloading the heavy workload to us. Businesses can move fast and break things without compromising on security fundamentals. They can use one of two prebuilt security engines for low latency and high intensity applications, or we'll design a personalized approach for highly complex business use cases.

APIs

1. Drop-in Model Router - Replace your existing API call with ours

2. Threat Analysis - Get a security assessment before making your request

Security Engines

Each security engine's sole focus is to analyze the user prompt and past message history to determine whether it's safe to make the request to the LLM (before the model can take actions).

• sav-micro - lightweight, low latency security

  • Highly fine-tuned and trained on recent prompt injection exploits

• sav-heavy - multi-layered protection for advanced applications

  • Larger models and more thorough threat analysis

• Personalized approach for a specific business need

What’s Next

If you read this far, thank you! Subscribe below for more updates, and feel free to email me at colin@savira.dev if you have any ideas or questions. If 100+ people sign up for the waitlist/newsletter, I'll commit to building the product, so please share this with others you think might be interested! I’ll be documenting everything related to building Savira, from technical deep dives on AI security challenges to early product demos and beta access opportunities. Check out my portfolio here if you want to learn more about me!