DeepKeep Reveals InkJect: New Visual Prompt Injection Vulnerability in AI Models

DeepKeep has identified a new visual prompt injection vulnerability named 'InkJect' that affects leading visual language models (VLMs) like OpenAI's GPT-5.2 and Anthropic's Claude Sonnet 4.6. This vulnerability allows malicious actors to embed hidden instructions within images, which VLMs process without detection, leading to unauthorized actions. The attack exploits a gap in existing security measures that focus on text-based prompt injections, leaving visual processing layers vulnerable. As enterprises increasingly adopt multimodal AI solutions, the risk associated with this vulnerability is expected to grow. DeepKeep's research highlights that this attack vector has received minimal academic attention, with only one prior paper addressing it. The hidden instructions can evade detection through techniques like white text on white backgrounds and distorted text perspectives. In tests, the vulnerability enabled an attacker to insert a backdoor into a website without the developer's knowledge.

Key Points: • InkJect vulnerability affects major VLMs, including OpenAI's GPT-5.2 and Claude Sonnet 4.6. • Malicious instructions can be embedded in images, bypassing existing security measures. • The vulnerability has received little academic attention, indicating a significant security gap.