You gave your AI agent access, but who is giving it orders?

Research by Gen Threat Labs showed that AI agents can be manipulated by malicious content encountered during normal operation, causing them to execute harmful commands without any software vulnerability being exploited. This risk arises because agents treat untrusted content as instructions. The Agent Trust Hub (ATH) developed by Gen Digital provides layered protections including scanning and verifying skills before installation, a curated marketplace of trusted skills, an AI-native VPN for secure agent connectivity, and Sage, a runtime safety layer that intercepts and evaluates potentially dangerous actions before execution. Sage uses multiple detection layers such as URL reputation and prompt-injection detection and is open source to allow extensibility. Additional tools include AI Data Checker for transparency on data usage and collaboration with AMTSO to establish testing standards for agent protection products. These measures aim to mitigate risks inherent in autonomous AI agents acting with user permissions.

The impact includes unauthorized execution of system commands, data exfiltration, and destructive file operations by AI agents manipulated through malicious inputs. This can lead to compromise of user data, credential exposure, and potential system damage. Since no software vulnerability is required, traditional patching does not address this risk. The threat extends to the integrity and trustworthiness of AI agents as they gain more autonomy and permissions, increasing the attack surface through malicious skills, plugins, or prompt injections. Additionally, malware like Warden Stealer targets AI assistant configuration files, further risking credential and data exposure.

Gen Digital provides the Agent Trust Hub as a layered security solution that includes skill verification and scanning before installation, a curated marketplace of trusted skills, and runtime safety controls via Sage that intercept and evaluate agent actions before execution. Secure connectivity is enforced through an AI-native VPN designed for agents. Users should leverage these tools to control what agents can access and monitor agent activity. Sage is open source, allowing integration and extension for additional safety. The AI Data Checker tool enhances transparency on data sharing. Organizations should follow the AI Agent Runtime Safety Standard (AARTS) and use products tested under AMTSO guidelines for reliable protection. Since this threat does not rely on software vulnerabilities, traditional patching is not applicable; instead, adopting these layered controls is essential.