I was playing around with prompt injection attacks on AI coding agents and kept thinking - there's literally no way to just test how secure your agent actually is. Like you can read all the papers about it but nobody built a tool that actually throws hundreds of attacks at your agent and gives you a score. The thing that bugged me most is that with AI agents, the data and the instructions are the same thing. Thats not something you can just fix with a patch, its how LLMs fundamentally work. So security isnt about making it bulletproof, its about measuring how much risk you actually have. I started with a simple open source CLI that runs extraction and injection probes against any LLM endpoint. Pretty straightforward. But then it kinda grew in directions I wasnt expecting. MCP servers became a huge thing. The ecosystem blew up to 3400+ servers and nobody was really checking them for security issues. So I built a scanner that does static analysis and toxic flow detection across all of them. That turned into its own whole product which I didnt plan for at all. Then I started testing actual AI agents like Claude Code, Gemini CLI, Codex in trap repos with hidden prompt injections and publishing the results as a leaderboard. That got way more attention than I thought it would. The biggest shift in my thinking was realizing a one time scan tool has a retention problem. People scan once and leave. So the approach evolved toward continuous stuff, MCP monitoring, security badges that server authors put in their READMEs, a dashboard for tracking security over time. Still figuring out the right balance honestly but thats where its heading.

AgentSeal

Break Your Agents Before Someone Else Does

Break Your Agents Before Someone Else Does