TLDR Tech / InfoSec
When the Source Attacks Back: Prompt Injection Is Coming for OSINT
Security researcher Dekens has demonstrated how adversaries can manipulate AI-assisted OSINT investigations by planting hidden prompt injections within digital documents. Through a dedicated training lab, Dekens highlights techniques like pushing CSS div elements off-screen, embedding white-on-white text in PDFs, and hiding instructions in unformatted spreadsheet cells. While not yet observed in active campaigns, these invisible artifacts successfully steer language models toward false narratives without altering the visually rendered file. To defend against poisoned sources, investigators should hash original documents and independently extract text using tools like Apache Tika or Tesseract to compare against the visual output. Furthermore, analysts must strictly isolate AI models in read-only environments to prevent untrusted inputs from executing unauthorized actions.
Read the full story