Skip to content

Community

Provael is built in the open by a solo maintainer, with the Embodied AI Security Top 10 meant to become a community-owned list. Here is where things happen and what to expect.

Where to talk

GitHub Discussions is the front door:

  • Announcements — releases, new studies, roadmap notes (maintainer-posted).
  • Top-10 RFC — propose or dispute a risk for the Embodied AI Security Top 10; see the RFC process.
  • Results — share a run, a transfer result, or a checkpoint you red-teamed. Numbers welcome — with their controls (ASR + 95% CI + benign FPR).
  • Q&A — usage questions, adapters, calibration, CI wiring.

For a security vulnerability in the tool, do not use Discussions — follow SECURITY.md (private advisory). For a bug or an over-claim, open an issue with the matching template.

What to expect

  • This is a solo, build-in-public project — best-effort response, usually within a few days.
  • Security reports are acknowledged within 3 business days (see SECURITY.md).
  • Every claim ships with its control. If you think a number is over-claimed, the evidence-defect issue is the fastest way to flag it.
  • Contributions need a DCO sign-off (git commit -s) — see CONTRIBUTING.md.

Build in public

Roadmap, findings, and honest nulls are published as they land — in the changelog, the findings, and the pre-registered studies. A cleanly-measured negative result is a first-class output here, not a failure to hide.

Where this project participates

Accurate as of 2 September 2026. The dates below are other people's deadlines and will go stale — check the linked source before relying on one.

Community indexes. Provael has open submissions to a number of community-maintained lists of embodied-AI and ML-security tooling. Those are other people's repositories on other people's schedules, so the honest statement is that the submissions are open, not that they have all landed.

Open calls this work is aimed at.

  • NIST AI Standards "Zero Drafts" — the initial public draft on public-facing AI documentation is open for comment; NIST will consider input received by 16 September 2026. See the project page.
  • SPAIS 2026 — The Science of Physical AI Safety, a CoRL workshop (spais-ws.org). Submissions close 1 October 2026. Its third question asks whether robot foundation models need evaluation techniques meaningfully different from both LLMs and classical robotics, which is the question this tool answers in one narrow way. Worth stating plainly: the call is scoped to interpretability, alignment, control and evaluation, it does not name red-teaming or hardware-in-the-loop as topics, and Provael has no hardware result of any kind — so it speaks to part of what that workshop is asking and not the rest.

Reproductions. If you maintain a benchmark, a leaderboard or an index and you want a reproduction run against it, open an issue. Reproductions of other people's numbers are welcome here and get published whichever way they come out — including when they fail to reproduce ours.