Today, we're introducing Project Horizon, a new research initiative from SuperAlign focused on building context-aware safety systems for modern AI pipelines.
Project Horizon reflects our view that the next generation of AI security will not be built on rigid filters alone, but on systems that can better distinguish real risk from benign activity in the flow of actual enterprise use.
As organizations move from experimenting with AI to operationalizing it, the safety challenge becomes more nuanced. Teams need safeguards that can identify sensitive disclosures, adversarial behavior, and policy-relevant misuse without interrupting legitimate work, overwhelming analysts with noise, or degrading product experience.
That is the problem space Project Horizon is designed to explore.
In this first phase, Project Horizon brings together two specialized research systems: Halo and Prism. Both are being evaluated as part of a broader safety architecture for AI workflows, and both are currently available only through internal testing and a limited research-preview program for selected participants.
What SuperAlign is Solving
AI systems now sit closer to customer data, internal knowledge, operational workflows, and security-sensitive actions than ever before.
That shift has created a new class of safety requirements: not just blocking obviously harmful content, but understanding whether an interaction represents actual exposure, actual misuse, or merely language that resembles risk on the surface.
In practice, many existing controls still fail in one of two ways. They either overreact to harmless inputs that match suspicious patterns, or they miss subtle, contextual, and obfuscated behavior that does not look dangerous at first glance.
For teams deploying AI in real environments, neither outcome is acceptable. False positives erode trust and break workflows. False negatives create real operational and security exposure.
Project Horizon is our answer to that tradeoff.
We believe safety systems must become more context-aware, more operationally usable, and more aligned with how AI is actually deployed inside enterprises.
Two Moderation Systems
Project Horizon currently consists of two moderation systems that together form the foundation for a context-aware AI safety layer for effective moderation, data leakage prevention, and threat detection.
Halo
Halo is a lightweight safety classification system designed to help teams evaluate moderation, screening, and triage workflows across AI applications.
At a high level, it is built to support the identification of unsafe or policy-sensitive interactions and to provide structured signals that can be used inside broader review and decision pipelines.
Our goal with Halo is not to create a standalone enforcement engine. It is to make safety workflows more practical to prototype, test, and iterate, especially in environments where latency, reproducibility, and operational simplicity matter.
Halo is being developed for use in layered systems where model outputs inform downstream review, policy logic, and human oversight rather than replace them. That design philosophy matters: useful safety infrastructure should support judgment, not pretend to eliminate the need for it.
Prism
Prism is a research system focused on identifying potential disclosures of sensitive information across AI interactions and adjacent workflows.
Its core purpose is to help distinguish between text that merely references sensitive formats or policy language and text that may actually contain sensitive disclosures requiring attention.
That distinction is critical in real-world environments.
Enterprise teams routinely deal with prompts, logs, support interactions, and structured payloads that can look risky at a glance while being operationally benign. At the same time, truly sensitive content may appear in incomplete, indirect, or obfuscated forms that simpler approaches fail to catch.
With Prism, we are exploring how context-aware detection can improve first-tier screening for safety, compliance, and moderation use cases without defaulting to blunt pattern matching alone.
The broader objective is straightforward: reduce unnecessary friction while improving the ability to surface the interactions that actually deserve review.
Releasing Research Preview
Project Horizon is not a public product release today. This is a research preview intended for internal evaluation and a limited group of selected participants, with access focused on testing, feedback, and workflow learning rather than broad availability.
We are intentionally taking this approach because safety systems need to be validated in realistic operating conditions.
They must be evaluated not only for model quality, but for how they behave inside actual workflows, how they affect analyst load, how they interact with policy logic, and how they support responsible decision-making over time.
We will subsequently release the underlying technical reports, datasets, or detailed detection taxonomies as the project matures towards general availability.
At this stage, our focus is on careful iteration, applied evaluation, and collaboration with teams that want to help shape the direction of context-aware AI safety systems.
Both Halo and Prism are being developed for human-in-the-loop environments and layered safety stacks, not as autonomous systems for final high-stakes decisions.
We see them as components of a broader architecture that includes policy controls, logging, review workflows, and operational governance.
Get Access
We're opening Project Horizon to a limited set of early participants. If your team is building AI products, trust and safety infrastructure, security operations, or compliance workflows, we'd like to hear from you.
- Request research preview access if you want to evaluate Halo or Prism in a controlled preview setting.
- Apply as a design partner if you want to work closely with SuperAlign on shaping the next phase of context-aware AI safety systems.
Project Horizon is where we are exploring what comes next for practical AI security: systems that are more precise, more usable, and better aligned with the realities of enterprise deployment.
This is only the beginning.