Provael

Provael is a model-agnostic command-line tool that attacks open vision-language-action robot policies in simulation and reports an Attack Success Rate.

It ships 42 attacks across 19 families, 17 adversarial and 2 benign controls, covering perception spoofing, instruction injection, action-space hijacking, authorization, backdoors, and cross-domain safety misalignment.

Every reported rate arrives with the things that make it interpretable: a matched benign control at the same task and seed, a 95 percent Wilson confidence interval, and a transfer-status label. Attacks the environment cannot support are skipped and reported as N/A rather than scored zero. Output is a report.json and a SARIF file tagged with EAI rule IDs.

On a clean container, it is 20 seconds from pip install to a written report. A Docker image is published for amd64 and arm64 if you would rather not install anything.

Apache-2.0. The full simulator path requires Linux, because LeRobot declares its LIBERO simulator as Linux-only. The core tool runs anywhere Python 3.12 does.