Open-Source AI Security Scanner Blind to Non-English Prompt Injection Attacks

A developer auditing an open-source AI agent framework discovered that its built-in prompt-injection scanner detects threats almost exclusively in English. Testing 42 variations of seven distinct attack types across six languages yielded only seven detections — every single one in English, with all 35 non-English variants going undetected. The scanner proved fully effective against technical, language-agnostic payloads such as hidden Unicode characters and malicious shell commands, regardless of the surrounding language. The researcher stressed this is not a regional issue: any attacker can bypass an English-language deployment simply by writing their malicious prompt in another language. The finding points to a specific, bounded gap — regex-based intent detection built around English phrasing — rather than a wholesale failure of the security tool.