Model Armor filter version history

This document records the release dates for Model Armor filter versions and the improvements included with each release. For information about configuring filter versions and lifecycle stages, see Set the filter version.

Version release history

The following table records the changes introduced with each filter version release.

Version number Release date Description
v4 2026-11-12 The prompt injection and jailbreak detection filter is updated with the following improvements:
  • Higher recall for indirect prompt injection (IPI) and tool calls.
  • Higher recall for conversational inputs at a lower false positive rate.
  • Higher recall for jailbreak attacks.
2026-09-18 Minor updates have been made across prompt injection and jailbreak detection, and responsible AI filters to address reported false positives.
v3 2026-07-20 The responsible AI safety filter for the hate speech, harassment, sexually explicit, and dangerous content categories has been upgraded to improve detection accuracy and reduce the rate of false positives.
2026-06-22 The prompt injection and jailbreak detection filter is available in asia-southeast1. This filter is updated to improve accuracy and reduce false positives.
2026-06-16 The prompt injection and jailbreak detection filter is available in asia-south1 and northamerica-northeast2.
2026-05-25 The prompt injection and jailbreak detection filter is updated and uses a new model, which together have the following benefits:
  • Significantly reduced false positives.
  • Improved prompt injection and jailbreak detection.
  • Enhanced mitigation against various attack vectors.