Hidden Image Text Slips Past LLM Safety Filters

Research presented in San Francisco shows image channels can carry prompt injections that text-only screening misses.

Updated on Nov. 7, 2026 in Cybersecurity

Hidden Image Text Slips Past LLM Safety Filters

At the OWASP Global AppSec USA conference in San Francisco on November 5, AI researcher Pavan Reddy of Automata LLC demonstrated that malicious prompts can be hidden inside images to attack large language models (LLMs). The technique, which Reddy calls FigStep, is a research demonstration rather than a deployed exploit, and Reddy also presented mitigation steps.

Why it matters

LLMs generally process text and images through separate channels, and many open-weight models screen only text for malicious prompts or dangerous outputs. Because images pass through as raw data without being checked, the image path becomes an unguarded attack channel for the same class of prompt injection.

The demonstration measured how often a model fulfilled a phishing-email request embedded in an image, with success ranging from 60% to 80% depending on the model. The source does not specify how many models were tested or give a baseline for comparison against text-based attacks.

The players

Pavan Reddy

AI researcher and engineer at Automata LLC who presented the FigStep image-based prompt injection research.

OWASP Global AppSec USA

The application security conference where the research was presented, held in San Francisco.

The details

Reddy uploaded an image containing text steps to an LLM along with a text prompt asking the model to help complete those steps. Many LLMs use anamorphic resizing, which downscales the image resolution and raises contrast, so hidden text becomes visible to the model even when it is nearly invisible to a human viewer. Adjusting contrast and overlaying another image can make the embedded text nearly invisible to people while the model still reads it.

Timeline

  1. Thursday, November 5, 2026: Reddy presented the image-based prompt injection research at the OWASP Global AppSec USA conference in San Francisco.

The Tech Race

Reddy's demonstration places image-channel prompt injection alongside the text-focused defenses that current model providers and open-weight releases rely on. The research shows that securing LLMs requires screening every input modality, not only text.

Developers and companies deploying LLM-based tools that accept image uploads face exposure to this attack path until image inputs are screened. Mitigations described at the conference include adding OCR to image processors, which defeats the image-based attack, though the source does not give a deployment timeline for such fixes.

The takeaway

Watch whether major open-weight model providers add image-channel screening or OCR to their pipelines in the coming releases. Developers can check whether their image processors read embedded text before passing images to the model.

Further reading

For broader context on how attackers target AI systems, see Cybersecurity coverage across the United States.

Source note: This article includes information reported by SC Media.