Skip to content- Sources: primary, discussion, discussion
- Summary: IBM published granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview on Hugging Face under Apache 2.0, with a chat template that renders every prompt part as a tagged LLMON block and wraps tool results as data. MACCUS reports IBM's figures: on the OPI benchmark attack success falls from 96.3% to 0.02% for the 8B model, on the agent-focused ASB benchmark from 52.8% to 3.1%, with direct and multi-turn attacks still succeeding far more often. IBM labels the models for evaluation only and warns against making them the only defense for an agent.
- Why it matters: Teams running small local models over email, documents or tool output get a testable injection defense, but only if they adopt the bundled template and keep outside content in separate fields.
- Background: Indirect prompt injection hides an instruction in content the model reads, such as a web page or tool result, so the model treats attacker text as a user command. Earlier defenses like Spotlighting mark untrusted text with ordinary delimiters, which attacker text can imitate. LLMON instead uses reserved tokens from the model's vocabulary that plain text cannot produce, and an exec block names which instruction runs on which inputs.