Smuggling Prompt Injections Using Zero-Width Characters to Bypass Filters
GiiTZzz · reddit · 2026-08-13
A developer discovered that inserting zero-width spaces (U+200B) into normal-looking prompts can successfully bypass phrase-matching LLM security filters.
This technique is completely invisible to both humans and standard detection mechanisms, allowing stealthy injection of malicious commands like "ignore previous instructions." The author shared a defensive approach: specifically checking for zero-width and Bidi-override characters while carefully excluding legitimate Unicode used in RTL scripts like Persian or Indic to prevent false positives.
More from coding & agent
- Claude Code CLI Update: Bundle Size +8%, Adds Plugin Marketplace & SSE Keepalive — ClaudeCodeLog · 2026-08-13
- Claude Code CLI Update: Introduces Plugin Marketplace and Streaming Fixes — ClaudeCodeLog · 2026-08-13
- Beyond Trivialities: How LLMs Enable Deeper CS Learning — generativist · 2026-08-13
- AI Engineer World's Fair Highlights: Memory and Continual Learning for Agents — Stefania_druga · 2026-08-13
- The AI Agent email dilemma: How to handle replies from external clients — JanJanJaJa · 2026-08-13
- Open Source MCP Sandbox Computer: Instant VMs Dedicated for AI — Ambitious-Prompt-975 · 2026-08-13