Vision-language-action models now steer robot arms, humanoids and mobile machines. Absentry tests how easily they can be hijacked through what they read and see, and keeps the robot inside its job when they are.
Normal run. The arm follows its task inside the movement space.
Classic robot security protects networks and firmware. A robot driven by a vision-language-action model can also be misled by the very inputs it is built to follow.
A replaced or extended instruction can send the robot after a different task, or past the one it was given.
A patch or object placed in the camera view can derail the policy without anyone touching the robot.
Once the model is misled, the motion it outputs leaves the space the task needs, and nothing in the model notices.
Two products that work on their own and together: find out how your robot's AI fails under attack, then keep it inside its task when it does.
Before deployment
An attack suite run against your robot's policy in simulation. You get a report of what breaks, how often, and by how much, with the protocol behind every number.
In the control loop
A runtime layer between the model and the motors. When the robot is pushed out of its job, the guard brings it to a controlled stop.
The EU Machinery Regulation (2023/1230) applies from 20 January 2027, and for the first time it names machinery whose safety functions rely on machine learning.
20.01.2027
Machines must not act beyond the task and movement space they were designed for.
Machines must hold up against reasonably foreseeable malicious attempts to corrupt them.
Makers have to show how they met these requirements, with test results to back it.
Summary for orientation only, not legal advice.
We run security evaluations the way a lab runs experiments, so the results hold up when a regulator, a customer or your own team pushes on them.
Models, tasks, seeds, attacks, metrics and the pass bar are fixed before anything runs.
Clean and attacked runs share tasks and random seeds, so each difference is measured, not guessed.
Defences are tuned on development attacks and scored once on attacks and models they have never seen.
Failures and limits go in the report next to the wins, with what each result does and does not cover.
If a vision-language-action model decides where your robot goes next, its failure modes are now part of your safety case.
General-purpose robots that take open-ended instructions in homes, shops and plants.
Pick-and-place and assembly cells moving from fixed programs to learned policies.
Ground vehicles and drones that act on what their cameras see, far from an operator.
We are running pilot evaluations with teams building robots on vision-language-action models.