One person builds and maintains Provael.
That person is Sattyam Jain, an agentic-AI architect and tech lead. He is not a security engineer by trade, and he has built security tooling for AI agents before this. Provael finds and reproduces robustness regressions in vision-language-action (VLA) checkpoints, in simulation.
An architect who builds his own security tooling
Sattyam has shipped production AI across fintech, AI-training, and GenAI startups, leading delivery on enterprise systems spanning insurance, healthcare, legal, geospatial, and industrial software. What he brings to security is the habit of building the layer himself: agent-airlock, a deny-by-default contract layer for AI-agent tool calls, and agent-audit-kit, a static scanner for MCP-connected agent pipelines that reports in SARIF.
He created pyAGI, an autonomous-agent Python framework, in 2023. It was later acquired by two independent builders. Their names appear here only with their written permission, which has not been asked for yet, so they are not named on this page.
His work centres on the parts of AI that are hardest to trust - agentic-AI governance, security, and observability - and on mapping AI systems to published frameworks (EU AI Act, NIST AI RMF, OWASP, ISO).
Why a robot policy needs its own red-team
Most AI-security tools test what a model says. A vision-language-action policy turns language and perception into motion, so its failure is a trajectory across a keep-out line rather than a toxic paragraph. Text-only tools cannot reach that action layer, and OWASP’s agentic lists target software agents, without physical actuation.
Provael measures that layer: an attack-success rate with a 95% confidence interval and a benign control, reproducible from a committed recipe, that you can file as evidence or use to gate a release. The one measured result is on /results, and the risks it maps to are in the Embodied AI Security Top 10.
Negative results, published
Provael publishes negative results as loudly as positive ones. When an attack does not transfer to a real policy, the site says so, in the same sentence as the number.
How it is made. Much of the codebase, and much of this site, was written with AI assistance (Claude Code); every published number, calibration and security-relevant path is human-reviewed before it ships, and every commit is signed off by the person who reviewed it. The product’s README and CONTRIBUTING say the same.
- Who
- Sattyam Jain
- Role
- Founder and maintainer
- Base
- India
- Entity
- Being incorporated (Ahmedabad, expected January 2027); until then contracting is with the individual
- License
- Apache-2.0
Follow the work.
The notes are dated write-ups of it: what was measured, and what it does and does not show.