Provael

Provael is an open-source red teaming and assurance tool for vision language action policies, the AI models that convert camera input and natural language instructions into physical robot actions. Whereas existing adversarial testing tools such as garak, PyRIT, and promptfoo evaluate the outputs of language models, Provael evaluates the physical behaviour these models produce when embodied in a robot. The tool runs adversarial attacks against a policy in simulation and measures the resulting attack success rate together with a statistical confidence interval and a benign control condition. Results are output as machine readable evidence, including SARIF, OSCAL, and CycloneDX ML BOM formats, intended for use in continuous integration pipelines or regulatory review. Findings are mapped to an independently authored taxonomy of embodied AI security risks and cross referenced to current regulatory frameworks, including the EU AI Act, the EU Machinery Regulation, ISO 10218, the NIST AI Risk Management Framework, IEC 62443, and the EU Cyber Resilience Act.

The core functionality of the tool, including all attack families, statistical scoring, evidence generation, and the continuous integration gate, is available at no cost under the Apache 2.0 licence. Paid services are limited to hosted evaluation runs on graphics processing hardware, signed leaderboard entries, and compliance documentation packages.