Search papers, labs, and topics across Lattice.
This paper introduces HarnessRisk, a comprehensive benchmark for evaluating the safety of agent harnesses across six operational phases, addressing the limitations of existing safety benchmarks that focus on isolated attack mechanisms. By analyzing 128 sandboxed cases, the authors reveal significant vulnerabilities, particularly in the Harness Configuration phase, where attack success rates can reach up to 80.9% despite high utility scores. The findings underscore the inadequacy of current safety measures, as even configurations that detect risks often fail to prevent successful attacks, highlighting the complexity of ensuring safety in multi-faceted operational environments.
HarnessRisk reveals that up to 80.9% of adversarial attacks can succeed in agent harnesses, even when risk detection is high.
Large language models are increasingly deployed through agent harnesses that manage tools, extensions, persistent state, permissions, and external actions. Existing safety benchmarks mainly target individual attack mechanisms or a limited subset of operational settings, making it difficult to compare how safety failures emerge across different harness responsibilities. We present HarnessRisk, a lifecycle oriented benchmark that organizes agent harness safety into six operational phases including Harness Configuration, Capability Extension, Runtime Operation, State Persistence, Action Control, and Incident Recovery. HarnessRisk contains 128 sandboxed cases, each pairing a benign user objective with an adversarial instruction embedded in an untrusted workflow artifact. We evaluate each trajectory using Utility, Attack Success Rate, Persistence, and Detection. Across three harnesses, six language models, and 14 model and harness configurations, attack success ranges from 12.6% to 80.9%, while Utility remains between 75.0% and 97.6%. Harness Configuration is the most vulnerable phase across all three harnesses, showing that attacks can succeed by altering security sensitive parameters within otherwise authorized workflows. We also find that explicit risk recognition does not reliably lead to safe action, as some configurations detect risks in more than 90% of runs while retaining substantial attack success. These results highlight the need to evaluate agent safety across multiple harness responsibilities and at the level of the deployed model and harness configuration.
