news.mlab.sh
Back to the feed
Add to case
threat-intel

New Agent Data Injection Attack Can Make AI Agents Misclick or Run Attacker Commands

High
Image: The Hacker News
Summary

Researchers at Seoul National University, the University of Illinois Urbana-Champaign, and Largosoft have discovered a new attack method called Agent Data Injection (ADI) that can manipulate AI agents by subtly corrupting the trusted data they rely on. Instead of directly injecting commands, attackers plant fake information – like a fabricated email sender or a forged GitHub comment – to trick agents into performing unintended actions, such as clicking ‘Buy Now’ on a product page or running attacker commands. The attack works by exploiting how agents process data, treating punctuation as structural markers rather than actual characters, allowing attackers to insert misleading information without triggering traditional defenses. While current defenses against prompt injection are effective, ADI bypasses them by manipulating the agent's internal understanding of trusted data, and the researchers have demonstrated successful attacks against several leading AI agents, including Claude, Gemini, and Copilot. OpenAI, Google, and Anthropic have acknowledged the vulnerability and requested a copy of the research.

Read the full article at The Hacker News
Add news.mlab.sh as a preferred source on Google

Summary written automatically in our own words from the original article, which belongs to its publisher and remains the reference. It may contain errors. Sources & data

Report an error
Confirmed errors are listed on /corrections.