A recent investigation by researcher Rehberger revealed significant vulnerabilities in Anthropic's Claude AI, which can execute hidden commands in documents. By embedding innocuous code, he bypassed Claude’s controls, initially rejected commands due to safety protocols. Despite reporting the leak on HackerOne, Anthropic later clarified its bug bounty program covers such data exfiltration. Security experts warn that many AI models, including OpenAI’s ChatGPT and Google Gemini, show similar weaknesses in handling malicious requests.