Behavioral hardening beyond detection and blocking
Prompt-injection detectors and filters can miss novel, obfuscated, or context-dependent attacks. AirForge trains the next layer: safe model behavior after untrusted tool, MCP, RAG, or web content has already reached the context.