Google released a development kit to help programmers build AI agents, but researchers have discovered a serious security weakness that allows one AI system to take control of another, according to The Register.
The flaw works through hidden malicious instructions buried in code libraries that multiple AI systems share. When one AI system pulls in this poisoned code, attackers can hijack it and control what that system does. Researchers created the first documented example of this attack in controlled tests, proving that AI-to-AI hijacking is now possible.
The attack exploits something called prompt injection. A prompt is an instruction you give to an AI. Prompt injection means sneaking hidden commands into information that an AI reads, forcing it to follow those secret orders instead of its intended purpose. While programmers have known about prompt injection for some time, this is the first case where one AI system weaponizes it against another.
Why this matters becomes clear when you consider how AI systems are deployed in real business environments. More companies are building systems where multiple AI agents work together, share information, and use common code libraries. If one agent can hijack another through malicious code, attackers gain a foothold into entire connected networks. A compromise of one seemingly harmless tool could cascade into control of many interconnected systems handling sensitive data, money, or decisions.
Google’s development kit was designed to make it faster and easier for programmers to create these agent systems. The security research shows the kit, and similar tools from other companies, may lack strong enough protections against AI-to-AI attacks. The flaw reveals a blind spot in how companies are currently thinking about AI security. Developers have long been trained to protect systems from human hackers, but few have considered how one AI system might be weaponized against another.
Right now, this remains a problem demonstrated in controlled research settings. But as more companies deploy AI agents that work together in production environments handling real operations, this kind of attack could become a practical threat. The researchers published their findings to push the industry to build security frameworks designed specifically for AI-to-AI interactions, not just the existing approaches built for protecting systems from human attackers.
Source: The Register
