As organizations move from experimenting with AI to operationalizing it, traditional safety filters alone are no longer enough. Project Horizon addresses the nuanced challenge of distinguishing real risk from benign activity — helping teams identify sensitive disclosures and potential misuse without disrupting legitimate work.
Introducing Halo and Prism
Project Horizon ships with two specialized moderation systems built for enterprise AI environments.
Halo
Halo is a lightweight safety classification system that helps teams evaluate moderation, screening, and triage workflows across AI applications. It identifies unsafe or policy-sensitive interactions and provides structured signals for downstream review and decision-making — without replacing human judgment.
Halo is ideal for teams that need practical, low-latency safety prototyping in production environments.
Prism
Prism is a research system focused on detecting potential disclosures of sensitive information across AI interactions. It goes beyond pattern matching to understand context — helping distinguish between text that merely references sensitive formats and text that may actually contain sensitive data requiring attention.
This reduces unnecessary friction while improving the ability to surface the interactions that actually deserve review.
Availability
Project Horizon is currently available as a research preview for select participants. This limited release is designed for organizations interested in testing, providing feedback, and helping shape the future of context-aware AI safety.
Both systems are built for human-in-the-loop environments and layered safety stacks — not as autonomous decision-makers. They integrate with existing policy controls, logging, review workflows, and operational governance frameworks.
What's Coming
As Project Horizon matures, we will release technical reports, datasets, and detailed detection taxonomies to the community. Our goal is to make context-aware AI safety accessible, practical, and aligned with how enterprises actually deploy AI.