Safety, alignment & governance
StepJack: Benchmarking Computer-Use Agent Safety Against Multi-Step Indirect Prompt Injection
StepJack benchmarks a multi-step injection threat: adversarial goals are split into innocuous-looking actions across pages in an agent’s navigation chain.
Why it matters: Evaluations limited to single-step injections may miss risks from instructions distributed across pages an agent visits. The reported aggregate also applies only to agents that can reliably follow the reference chain.