Zscaler finds autonomous AI agents vulnerable to indirect prompt injection fraud schemes

According to CSO Online, Zscaler tested major LLMs and found autonomous AI agents falling victim to indirect prompt injection (IPI) traps and fraud schemes that would deceive few humans. Vulnerable models included Llama3-3-70b-instruct, Llama3-2-90b-instruct, Gemini-3-flash, and Gemini-2.5-pro, while Llama4-maverick, Gemini-3.1-pro performed as safe. The testing demonstrates that high-end enterprise agents can be compromised by schemes designed to exploit their reasoning without human visibility.