Prove it. Prevail.

The open-source red-team & assurance layer for physical AI.

Robots and humanoids increasingly run on VLA (vision-language-action) models — a single neural network that turns a camera image and an instruction into motor commands. That "robot brain" can be attacked much like an LLM is jailbroken, except the failure now moves a real arm. Almost no one is security-testing this layer yet.

Provael attacks VLA policies in simulation, measures an Attack Success Rate (ASR), and is building toward proving a policy is safe — against the standards regulators and insurers are beginning to require (ISO 10218:2025, the EU AI Act / Machinery Regulation).

➡️ provael — the red-team harness. Model-agnostic, CPU-first, Apache-2.0. Perturbs the instructions and observations a VLA receives and reports how often it's driven into an unsafe state.

One honest early result: a simple instruction reframing diverted a real SmolVLA × LIBERO policy 100% of the time on a pick-and-place task — while the benign baseline stayed at 0%. Visual/scene-text attacks didn't move it (0%). Early, reproducible, and we say exactly what does and doesn't work.

The robot security lifecycle, in the open: attack (red-team VLA policies) → prove (assurance reports mapped to the standards) → guard (runtime checks). Built in public — wins and dead ends.

🔗 Repo: github.com/provael/provael · 🐦 X: @getprovael · 🌐 provael.com (coming soon) · ✉️ getprovael@gmail.com

_{Founded by Sattyam Jain — GenAI architect, agentic-AI security (agent-audit-kit, agent-airlock).}