Introduction As large language models (LLMs) and deep learning technologies are widely applied in the security field, traditional webshell detection methods based on feature codes, abstract syntax trees (AST), or sandbox behavior analysis are gradually evolving towards AI-based semantic analysis. AI models can understand the logical intent of code, thus identifying malicious scripts that have been obfuscated in complex ways. However, AI models themselves have an inherent weakness: when processing code, they often reason about code logic alongside "non-executable content" such as comments and metadata in the same semantic space. This creates an opportunity for prompt injection (\n\n---\n📊 Market Data: BTC $71,816.19 (-3.59%)\n#Crypto #Bitcoin #Binance