The original headline is: "I red-teamed my own LLM security gateway in four passes. Here's every gap I found."

Short summary

A developer red-teamed their own LLM security gateway across four sessions, documenting every gap found between attack and benign corpora. Key findings include ASCII smuggling via Unicode tag characters, ChatML control-token injection, and overly broad exfiltration rules causing false positives on legitimate security questions. The author honestly documents five gaps that regex alone cannot solve, emphasizing that pattern detection raises attack costs but doesn't end the game.

• •Four-pass red-team of an LLM security gateway with attack and benign corpora
• •Key gaps: ASCII smuggling, ChatML token injection, false positives on defensive queries
• •Author transparently documents regex limitations and five unsolvable gaps

Generated with AI, which can make mistakes.