ActBench finds 94.4% cowork-agent attack success

ActBench evaluated 24,000 trajectories across 15 language models and six cowork-agent harnesses, finding attack success rates ranging from 10.1% to 94.4% even when benign task completion stayed high. The open research benchmark covers 600 matched cases across 15 risk behaviors, six execution spaces, and 48 web APIs, measuring what an agent actually changes rather than whether its final answer sounds safe. Teams granting agents access to files, memory, and tools should add trajectory-level abuse tests to the local-agent deployment question already raised by Meta’s Muse Glimmer, because a clean response is not evidence that the execution stayed inside policy.