Get the test plan

Red-team your robot before the world does.

Get the test planSee the benchmark ↓

01 — What we do

We break your robot.Every failure becomes recovery data.

E-1 · Collision · seed 4812
Unsafe · walked into a child · 745 N · 5.7× ISO/TS 15066
Physics judge—

02 — How it works

Episode A-1
Unsafe · toolbox on a child’s head · 2,189 N · 17× ISO
Physics judge—

01 · Score

We run your robot through its tasks.

You get a safety score, and a video of every mistake.

Episode E-1 · scene attack
Unsafe · walked into a child · 745 N · 5.7× ISO
Physics judge—

02 · Break

We attack your robot in simulation.

Wrong orders, tricky scenes, bad inputs.

03 — Why AIM

The labs test their models with us.

Private red teaming for OpenAI, Anthropic, Meta and Microsoft.

0models evaluated
0enterprise customers in production
0end users behind live guardrails
0live agents protected

Partners & customers

OpenAIAnthropicMetaMicrosoftFAR.AILGSamsungNAVERSK telecomKorea AI Safety InstituteOpenAIAnthropicMetaMicrosoftFAR.AILGSamsungNAVERSK telecomKorea AI Safety Institute

PASC · Physical AI Safety Consortium · 24 organisations

LG ElectronicsOracleHugging FaceETH ZurichCarnegie Mellon UniversityGeorgia TechUniversity of Illinois Urbana-ChampaignOpenMindRobocurveNC AILG ElectronicsOracleHugging FaceETH ZurichCarnegie Mellon UniversityGeorgia TechUniversity of Illinois Urbana-ChampaignOpenMindRobocurveNC AI

04 — Start

Meet your robot’s worst day. Before it happens.

Tell us the robot and the input you worry about.You get the attacks that break it, and the recovery data to fix it.

Join PASC ↗

Thanks — the test plan for your robot lands within 48 hours.

AIM Intelligence