CIBER: A Comprehensive Benchmark for Security Evaluation of Code Interpreter Agents

The researchers propose CIBER, an automated benchmark for evaluating the security of code interpreter agents, combining dynamic attack generation, isolated sandboxes, and state-aware evaluation. The benchmark covers four categories of adversarial attacks: direct and indirect prompt injection, memory poisoning, and prompt-based backdoors, and evaluates six foundation models on two agents, OpenInterpreter and OpenCodeInterpreter. Results show that attacks disguised in natural language achieve a 14.1% higher success rate than explicit code snippets, and that agents defend well against explicit threats but fail against implicit semantic risks, which the authors call security polarization.