BATTLING SHADOW AI: PROMPT INJECTION FOR THE GOOD

Summary from the TLDR InfoSec newsletter

Researchers developed a defensive use of prompt injection to combat Shadow AI risks by embedding hidden warnings in corporate documents that trigger alerts when users upload sensitive files to unapproved AI tools like ChatGPT or DeepSeek, successfully displaying disclaimers and, in some cases, blocking processing entirely. The team built an open-source prototype tool called "Prompt Injection for the Good" to test effectiveness across multiple LLMs and file types, finding most models honor carefully-phrased warnings in .docx, .pdf, and .eml formats. It failed to detect prompts in heavily obfuscated text and OCR-based tools. Security professionals should experiment with embedding defensive prompt injections in corporate document templates and sensitivity labels as a user awareness mechanism for Shadow AI. This creative approach has limitations, including inconsistent LLM behavior, potential conflicts with vendor anti-injection defenses, and the dual-use risk of the technique itself.