Runtime guard for computer-use agents reports a 16.5-point accuracy gain
single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC
What this means for you
Nothing to act on yet: there is no model to download, no API and no date. If you run computer-use agents in production, the interesting claim is the method, not the weights — scoring whole decisions instead of tokens is checkable, but only once the code and evaluation artifacts actually appear.
Researchers have released HazardAuditor, a safety classifier for computer-use agents — software that drives a browser, terminal or file system on a user's behalf. Existing guards score prompts and replies; this one is trained on runtime behaviour, running Claude Code, Codex, Hermes and OpenClaw in controlled environments and normalising their traces into one event format. A training method the authors call GuardPO optimises whole safety decisions rather than individual tokens, so longer explanations do not dominate the update. They report accuracy gains of up to 16.5 percentage points over the strongest prior guard, on unnamed benchmarks with no evaluation date. Code and models are promised, not yet posted.
Key facts
- ·HazardAuditor runs four agent frameworks — Claude Code, Codex, Hermes and OpenClaw — in controlled environments and normalises their interactions into a canonical event representation. source
- ·The authors report accuracy improvements of up to 16.5 percentage points over the strongest prior guard model; the benchmarks are not named in the abstract. source
- ·No evaluation date is given for the reported accuracy figures. source
- ·GuardPO converts deterministic safety outcomes into sequence-level advantages and normalises rationale and verdict regions, making the safety decision the unit of optimisation. source
- ·Code, models and evaluation artifacts are described as forthcoming at yunhao-feng.github.io/HazardAuditor/, not yet available. source
What the sources say
- Hugging Face Daily Papers (research) — Abstract describing an execution-grounded safety classifier and a training objective across four agent frameworks.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersHazardAuditor: From Executable Threats to Safer Computer-Use Agents2026-09-13