Offer: sim-only red-team (instruction-redirection) eval of MolmoAct2
#1
by Sattyam - opened
Hi team,
I run provael, an open-source red-team + assurance layer for VLA policies. It perturbs the instructions and observations a policy gets in simulation and reports an attack-success rate with 95% confidence intervals and a benign false-positive control, mapped to an embodied-AI security Top 10 and the arriving robot-safety rules. MolmoAct2 being fully open makes it a strong candidate to red-team in the open.
I'd like to run our action-space / instruction-redirection transfer test on a released checkpoint and publish the result responsibly: sim-only, transfer-tested on held-out seeds, no "first" claims, with the canonicalization defense noted next to any finding. Two questions before I start:
- Is there a preferred checkpoint and LIBERO/eval config you'd want me to test against?
- Any responsible-disclosure preference if something transfers?
Happy to share the method first.
Sattyam