Preprint: an instruction hidden in an image can become visible to a model when the image is heavily downscaled
According to a preprint, a text instruction can be hidden in an image so that four closed vision-language models read it in more than half of tests, but only once the image had been downscaled to 64 × 64 pixels. With milder downscaling, the effect weakened sharply.
Researchers at Nanyang Technological University and DSO National Laboratories describe a phenomenon they call Scale-Induced Semantic Shift (SISS) in a preprint published on arXiv on 12 November. The study has not been peer reviewed.
Systems often downscale or compress an image before a vision-language model processes it. The model may therefore not see the same image as the user.
Tests and results
The researchers fine-tuned an image generator on images they selected solely on the basis of feedback from the open-weight Qwen2.5-VL model. The test images were frozen before the target models saw them. An attack succeeded if a model did not read the instruction in the 512 × 512 pixel image but did read it in the version downscaled to 64 × 64 pixels. Gemini 3 Pro acted as judge and decided from the text response alone whether the instruction had been found. It did not see the image.
Attack success rate on 300 test images:
| Model | SISS | QR-Monster (baseline) |
|---|---|---|
| GPT-4o | 60.33% | 35.33% |
| Claude Sonnet 4.6 | 57.67% | 49.33% |
| Qwen3.5-Plus-02-15 | 62.67% | 47.67% |
| Gemini-3-Flash | 54.33% | 32.00% |
| Average | 58.75% | 41.08% |
Source: authors’ preprint, third refinement round, downscaling from 512 to 64 pixels.
The figures do not show that the attack works at ordinary image sizes. The authors said 64 pixels was a deliberately aggressive test setting, not an assumption about how deployed models work. With the same images, the average success rate was 17.68 per cent at 128 pixels, 6.85 per cent at 224 pixels and 5.60 per cent at 336 pixels.
Small effect on agents
In a web-agent experiment, the images increased detection of the hidden instruction. For Gemini-3-Flash, the rate rose from 37.33 per cent with control images to 67.67 per cent, and for GPT-4o from 12.67 per cent to 30.33 per cent. However, the agents took harmful actions in only 4.00 and 3.67 per cent of cases respectively, compared with 2.33 and 1.67 per cent with the control images. The authors said the experiment shows that images can influence an agent, but not that they can reliably hijack one. The experiment involved only two agents.
Generic image filters offered no protection
In the defence experiment, rescaling, JPEG compression, filtering and random masking produced no statistically significant change in the success rate. The authors’ ScaleGuard detector compares the full-size image with nine downscaled versions and rejects the image if text appears only in the small ones. It cut the success rate from 61.4 per cent to 24.3 per cent. When the attacker refined images for one round against the detector, the rate fell from 42.5 per cent to 15.0 per cent.
The detector rejected none of the clean control images in samples of 70 and 40 images. The authors said samples this small do not show that false alarms are rare. Nor did they measure the detector’s latency, even though it checks ten versions of every image. The defence experiment used a different dataset and judge and targeted only the Qwen2.5-VL-7B model, so its figures cannot be compared with the main results.