The updated model introduces a 1-million-token context window, allowing security teams to ingest extensive investigation histories and complex repository structures without fragmenting evidence. By prioritizing reproducible results over raw alert volume, the system attempts to bridge the gap between theoretical weaknesses and validated risks. In recent controlled demonstrations, the platform successfully executed tasks ranging from SQL injection bypasses and business-logic manipulation to the autonomous generation and verification of security patches.

Continuum AI reported that the model follows the success of its predecessor, which previously achieved over 98% pass@1 on CyberGym. Beyond cloud-based access via the OrcaRouter gateway, the company intends to expand its reach by releasing a specialized open-weight model next week. This upcoming release is designed for local deployment, providing organizations with direct control over sensitive security workloads while maintaining the reasoning capabilities of the frontier model.