Policy demoCatch a poisoned skill
Skills from outside marketplaces can contain hidden instructions, such as quietly sending files off the machine. Qpoint uses a System One model to review any skill that isn’t already trusted before the agent loads it. In the demo, a skill that sends out a credentials file is caught, and normal skills still load as usual.
Gaurav Mahajan
October 28, 2026
About this demo
Skills from external marketplaces can look harmless on the surface and still contain instructions you’d never approve. In this demo, we build a skill that quietly sends a credentials file to an outside server and tells the model not to mention it.
Our Qpoint policy combines a static list of trusted skills with a dynamic check for everything else. Any skill not on the trusted list is evaluated by a judge model before it’s loaded: does this skill tell the agent to do something the user wouldn’t expect, like sending data off the machine, reading credentials, or running downloaded code?
The trusted skill loads. Safe, untrusted skills load too, with no false positives. The poisoned skill is denied before it ever reaches the model.
See Qpoint in action in your environment
Request access and we'll get you set up.