Cybersecurity researchers at Tenet Security describe an attack they call “Agentjacking,” which they say can hijack AI coding agents so they execute arbitrary code on a developer’s machine. The researchers report that the method does not require malware, stolen credentials, or any direct breach of the target system. Instead, the attack relies on manipulating the inputs that lead a developer’s AI coding agent to take action.
According to the published accounts, Agentjacking can be triggered by submitting a crafted, fake bug or error report. The reports are created using Sentry, an open-source error-tracking and monitoring platform. When a developer routes the issue to their AI coding agent—such as by asking the agent to analyze the report or fix the problem—the agent may be induced to run malicious instructions embedded in the crafted report.
The described technique is presented as a new class of abuse that turns the coding agent itself into an attack mechanism. The reports emphasize that the attack focuses on agent behavior and prompt/issue handling rather than exploiting traditional vulnerabilities in operating systems or applications. The full details and broader implications are discussed as part of the disclosure by Tenet Security.
