LaunchPolicyNovember 14, 2026

IBM Granite preview models resist indirect prompt injection

Read original source →maccus.fr

IBM released granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview (3.7B and 8.8B parameters), trained via its LLMON labeling system to separate user instructions from document content. On a public attack benchmark, injections succeeded 96-100% against standard Granite 4.1 but at most 0.1% against the previews.

1 source

More stories today

Open the live feed