Concerns Over Prompt Injection Attacks Through Document Ingestion in AI Security

The post discusses concerns about AI security, particularly vulnerabilities in document-based attack vectors. The author notes that while LLM outputs are often secured, the document input layer remains unprotected. They highlight that standard text parsers and antivirus tools may miss threats in documents like PDFs, which LLMs can still process. The poster seeks input from others who have encountered this issue in production.