A monitor that reads the model, not only its words.
Most safety checks judge what a model writes. Ulterior Safety’s detectors read what the model is doing inside while it writes, so they don’t depend on the model’s own safety training.
Ghost
Watches a model’s internal activity as it answers and flags harmful requests.
- Reads internal activity, not only the finished text
- Flags harmful requests before the reply is finished
- Keeps working when built-in safety has been removed
95.8% on JailbreakBench and 95.5% on XSTest at 1% false alarms. See the results
Reading activity, not text.
A request goes into an open model. As the model processes it, Ghost reads the model’s internal activity. It watches the internal signals, not the output text, so it can flag a harmful request before the reply is finished.
How it reads that activity isn’t published.
Wraith
The next generation of our detectors. It will be scored on the same held-out tests, and the results published on the Research page.
Brain Lab
The internal tool we use to explore what happens inside a model while it answers. It supports the research behind the detectors and isn’t offered publicly.
For teams that can’t rely on the model’s own safety.
AI labs
Evaluating and red-teaming open models, including copies whose safety training has been altered.
Teams deploying open models
Running models on their own infrastructure, where an independent check on harmful requests helps.
Evaluators and institutes
Groups that need a signal that doesn’t depend on what a model chooses to say.
Ghost is at research stage. There’s no public API or product release yet; a deployable monitor for self-hosted and open models is on the roadmap.
Talk to Ulterior Safety.
Investors, AI labs, teams deploying open models, and press. Messages go straight to the founder.