ARTFEED — Contemporary Art Intelligence

HarnessSafe: Benchmarking Persistent-Carrier Safety in Agent Harnesses

ai-technology · 2026-08-10

A new benchmark named HarnessSafe has been introduced to evaluate safety across persistent carriers in agent harnesses. The benchmark comprises 328 executable cases spanning seven persistent-carrier families and is designed to assess mainstream agent harnesses. Each case is structured as a Persistent-Risk Lifecycle, tracing attacker influence from initial entry through persistence across carriers and system boundaries, culminating in a benign trigger and observable violation. The evaluation method is multi-stage and trace-based, relying on observable execution evidence to identify risk propagation. This work addresses the delayed safety risks posed by persistent state in agent harnesses, which existing benchmarks often overlook by focusing on a few carriers or harnesses. The paper is available on arXiv under the identifier 2608.06984.

Key facts

  • HarnessSafe is a new benchmark for evaluating safety in agent harnesses.
  • It includes 328 executable cases across seven persistent-carrier families.
  • The benchmark covers most mainstream agent harnesses.
  • Each case is specified as a Persistent-Risk Lifecycle.
  • The evaluation is multi-stage and trace-based.
  • It uses observable execution evidence to assess risk propagation.
  • The work addresses delayed safety risks from persistent state.
  • The paper is available on arXiv (2608.06984).

Entities

Institutions

  • arXiv

Sources