(Industries) Physical AI
Kyde detects when an agent that controls a machine has been compromised.
Behavioral scoring for AI agents.
For teams that run robots and machines with vision-language-action models or agents.
Where an agent that controls a machine goes wrong.
(Hijack)
A line of text at the start of a run can take over the actions of a vision-language-action model.
(Backdoor)
A trigger hidden in the training data can pass normal tests and activate later.
(Drift)
A model can drift after an update or in a new environment.
Agents are starting to control machines through common interfaces.
- (Today)Most machine agents today are hard-coded.
- (Standards)With emerging standards such as Anthropic's Model Hardware Standard, in research preview since August 2026, agents control robots and machines through common interfaces. Each command becomes a tool call that can be recorded.
- (Kyde)Kyde learns which devices and value ranges belong to a task.
- arXiv:2506.03350Jones et al., 2025. A text applied once at the start of a run reaches much of a VLA model's action space, in simulation.
- arXiv:2510.10932Xu et al., DropVLA, 2025. A backdoor in the training data succeeds in about 99% of simulated runs, and in 20% of 200 trials on a real robot arm.
(Where Kyde sits)
Kyde sits above the safety controller and does not replace it.
A picking arm leaves its usual range.
Nothing in the format of the command is unusual.
(What happened)
- 06:10 A picking arm starts its shift, controlled by a vision-language-action model.
- 06:10 The task text it receives carries one added line.
- 06:11 The arm is sent to a position outside its work cell's usual range, at twice its usual speed.
(What Kyde did)
In the setup we are building with design partners, the command arrives as a tool call. Position and speed are outside the ranges Kyde has learned for this task. Kyde holds the command and alerts the operator. The safety controller stays in charge of the machine.
*Illustrative. Times and numbers are an example, not a customer's.
Kyde compares each run of an agent with the agent's normal behavior, using frequencies and statistics.
No language model is involved in the decision. The same record always produces the same score, and anyone with the record can recompute it.
Language models may be used later to help explain a finding. They do not make the decision. When an agent deviates, a person decides what happens next.
How to start
As a design partner, free of charge.
-
(Step 1) Connect
We connect to one machine agent you run, in a test cell or a pilot. Read only, the safety controller stays in charge.
-
(Step 2) Two weeks
Kyde learns the agent's normal behavior and shows you each run that deviates from it.
-
(Who you talk to)
Joerg Radehaus, co-founder, product · joerg@kyde.com Jürgen Eckel, co-founder, technology · juergen@kyde.com
*We work with a small number of design partners at a time.
(What is measured)
Measured so far on software agents. Physical AI is our next step, with design partners.
(What Kyde does not see)
Kyde does not see the machine's sensors or its safety controller. It sees the commands an agent sends, as tool calls.