Project Horizon
Context-aware safety that keeps models helpful while catching misuse and protecting sensitive data.
The gap in modern guardrails. Rigid filters fail when nuance enters the loop.
Safety frameworks from leading labs often rely on binary filters. They block obvious harms, but fail when nuance, teaching or multi-step context comes in.
The refusal problem
Excessive alignment causes a collapse in usefulness: models refuse benign or teaching queries because they cannot read intent.
Context blindness
Keyword-based PII and PHI detection struggles when sensitive data is buried in complex reasoning chains or emerges by combining prompts, leaving a hidden leakage risk.
Safety as a science, not a filter. Two specialized models that work alongside large language models.
Halo
Threat classificationPreviewThe primary reasoning engine for safety alignment. Instead of looking for keywords, Halo analyzes the intent and potential impact of a prompt.
- Adversarial robustness that detects jailbreak attempts and prompt-injection techniques that slip past traditional filters.
- A granular taxonomy that returns a multi-dimensional threat vector instead of a simple safe or unsafe verdict.
Prism
PII, PHI and PCI detectionPreviewA high-precision detection layer for sensitive data across regulatory domains, without breaking the reasoning operators rely on.
- Deep context awareness that tells random strings from structured PCI, PHI or financial identifiers in real medical and financial workflows.
- Dynamic redaction so models keep reasoning over anonymized data while compliance boundaries hold.
Human-in-the-loop evaluation. Safety can’t be solved in a vacuum.
Horizon is a live testing ground that relies on diverse, adversarial testing from a global community.
Compare side by side
Evaluators compare model behaviour under different safety configurations.
Vote to find the line
Structured votes define the Horizon Line: where a model turns from helpful to hazardous.
Share the benchmark
Data gathered in the preview feeds an open-source safety benchmark for the industry.
Capability meets safety. A decoupled mediation layer instead of over-tuned weights.
Safety logic runs in the Horizon layer between the prompt and the model, so the model’s core weights are never over-tuned and its performance holds.
Arrives
A request heads for the model.
Halo
Checks for adversarial intent before the model acts.
Prism
Scans for leaks across PII, PHI and PCI.
Answers
With its usefulness intact.
Phase one: research preview. Proving safety doesn’t have to cost intelligence.
Our thesis: context-aware mediation can stay helpful, catch adversarial misuse and hold strict privacy boundaries without collapsing what the model can do.
Open research
Future open-weight releases of Halo and Prism to speed up specialized data-protection research.
Enterprise policy engines
Real-time policy customization for defining safety boundaries dynamically.
Multimodal safety
Extending Halo’s reasoning to multimodal and agentic workflows.