The images behind the story
IBM’s Granite LLMON previews separate instructions from data. That is not a security guarantee.
Analysis: IBM’s experimental 3B and 8B models introduce a structured prompt format. Here is what the evidence supports and how to evaluate the boundary in your own application.
Article by GPT · Read the full story →
The editorial illustration
AI-generated editorial illustration of instruction and data separation; not a software screenshot or evidence of security performance.
- Image model
- openai/gpt-5-image-mini
- Why this image
- Separate instruction card and document stack illustrate the learned instruction/data distinction without presenting it as a guaranteed barrier.
Read the full image prompt
Editorial illustration about separating instructions from untrusted data: a precise brass sorting frame on a charcoal tabletop holds one clean ivory instruction card above a separate stack of textured green document cards. A thin amber thread from the document stack curls toward the upper card but stops at a visibly incomplete divider, suggesting an imperfect learned boundary rather than an impenetrable shield. Tactile paper sculpture, dramatic side lighting, simple bold composition, landscape 3:2. No typography, letters, numbers, logos, people, robots, padlocks, shields or software interface. Imagined conceptual illustration, not a benchmark diagram.
Produced by Maker Frontier · Image use & permissions
The image-generation model credited above made this illustration. It may differ from the model discussed in the article.
Code-rendered visual summary; AI-designed layout. Text comes from the article headline and summary.
View this image and its source →Produced by Maker Frontier · Image use & permissions