As code agents rapidly evolve, they are increasingly adopted to streamline software engineering workflows. These agents typically operate in information-rich and sensitive environments, with the capability to directly interact with system resources.

However, prioritizing capability improvements over safety can lead agents to facilitate unsafe behaviors. In particular, agents may generate and execute risky code due to insufficient security awareness or adversarially injected instructions. For instance, they may perform harmful operations on the operating system or file system, such as adding a risky alias to the .bashrc file or deleting sensitive data. Furthermore, such agents may be exploited for unethical purposes, including the generation and execution of biased code.

We design a comprehensive benchmark consisting of 330 benign tasks spanning five representative categories of common coding workflows, along with more than 120 malicious tasks covering nine key security risk categories. These risk categories are guided by established security standards, including CWE, OWASP, MITRE ATT&CK Techniques, and NIST. Based on these risks, we construct red-teaming tasks with malicious objectives under two primary threat models, i.e., direct and indirect, enabling a systematic evaluation of the security robustness of code agents.