Security Researcher Discloses Data Exfiltration Flaw in Anthropic's Claude

Indirect Prompt Injection Attack

A security researcher, Johann Rehberger, disclosed a vulnerability in Anthropic's Claude AI that can be abused for data exfiltration. The attack does not exploit a traditional bug but rather the model's intended functionality. The attack method is known as "indirect prompt injection." This occurs when a third party hides malicious commands inside a document, such as a PDF or text file, that a user then uploads to the AI for a legitimate purpose like summarization. This vulnerability is part of a growing class of AI-specific risks. A separate report from Artificial Intelligence News warned that …