Threat model

How Ovrin handles hostile documents: prompt injection, zip bombs, parser exploitation and exfiltration, with the residual risk stated for each.

Document extraction systems face risks from malformed files, hidden text, prompt-like content, and extraction traps. Ovrin addresses this by giving the system finite limits and making validation part of the result.

Core principles

  • treat input as untrusted
  • validate the output against the schema
  • detect values that are not grounded in the document
  • prefer explicit review over silent acceptance