We built glc PromptGuard because production AI agents need more than a keyword filter. Untrusted text can arrive through users, retrieved documents, and tool outputs, so the service evaluates each new text delta with eight complementary detection layers and returns a clear verdict plus structured policy hints. The current version was evaluated through millions of end-to-end runs built from several hundred thousand multilingual attack and benign examples. In our latest live full-pack test, version 0.3.25 passed 1001 of 1004 cases; the three misses were the same known Hebrew scientific-abstract hard negative, classified as a false positive. We publish the methodology and versioned summary for security reviewers. Integration is intentionally simple through remote MCP or REST. Every human or agent account receives 100 free credits; typical checks use 1–2 credits.

glc PromptGuard

Stop prompt injection before it reaches your AI agent

Stop prompt injection before it reaches your AI agent