.jpg)
Artificial intelligence agents are transforming the way companies automate tasks, from web browsing to code execution. However, alongside their widespread adoption, new attack surfaces emerge that jeopardize system integrity. A recent study has identified a previously underexplored threat category: data injection attacks on agents (DIA). Unlike well-known indirect prompt injections, DIAs manipulate critical metadata or context formats that the agent interprets as trusted data, causing unintended actions without the user noticing. This type of vulnerability affects both web agents and coding assistants, enabling everything from arbitrary clicks on interfaces to remote code execution and supply chain attacks.
The root of the problem lies in the fact that current agents do not properly separate trusted data from attacker-controlled data. While traditional defenses focus on filtering malicious instructions, DIAs go unnoticed by disguising malicious information as a legitimate part of the agent's context. For example, a web agent could receive an altered resource identifier that leads it to interact with a fraudulent site, or a code agent could execute commands based on manipulated tool metadata. The lack of data isolation is, in essence, an omission of a fundamental security principle.
For companies that have already implemented AI agents in their workflows, this finding underscores the urgency of adopting a comprehensive cybersecurity approach. It is not enough to protect the prompt or the model's output; it is necessary to audit how context data is managed, validate sources, and apply integrity controls in every interaction. At Q2BSTUDIO we offer artificial intelligence services for companies that integrate good security practices from the design phase, helping to identify and mitigate vulnerabilities in agent systems. Furthermore, our experience in cybersecurity and pentesting allows us to assess the robustness of these solutions against emerging threats such as DIAs.
The evolution of attacks in the AI ecosystem requires organizations to combine custom applications with rigorous security policies. From custom software development to the implementation of AWS and Azure cloud services, each technological layer must consider the separation of trusted and untrusted data. Likewise, business intelligence tools such as Power BI can benefit from these practices when integrating agents that process external data. At Q2BSTUDIO we work with multidisciplinary teams to design secure, scalable architectures aligned with the most demanding standards, ensuring that AI innovation does not compromise business integrity.
