Submit ProductSign in
A family of custom-trained security models for coding agents
Grounded in available product and source data
Arcjet's new AI security models are built for a single job: reading agent traffic and deciding what looks like an attack. The first of the family, a general purpose model named Himalia 1, now runs the prompt injection detection that every SDK call can use.
Specialist versions are heading for situations general models handle badly, covering coding agents and autonomous security agents. The team trained its own models because public datasets disagree on what counts as an attack, overlap heavily, and rarely allow commercial use.
