An independent AI safety research company.
Founder-led, self-funded, and focused on one question: what is the model doing inside?
Why the company exists.
Checking what an AI says isn’t enough. Models can be jailbroken, have their safety training removed, or mislead the people using them, and output filters miss what the model is doing internally.
Ulterior Safety looks inside. It reads the model’s internal activity as it answers, to catch harmful behaviour before it reaches the output, including on models whose built-in safety has been removed.
The longer-term question is deception: can the same approach catch a model that knowingly misleads, hides what it is doing, or underperforms on purpose? That’s an early-stage research direction, not a capability.
Jasper Jensen
Jasper founded Ulterior Safety and built its whole research pipeline himself: running models, recording their internal activity at scale, and training and evaluating the detectors, from the first prototype to Ghost P.
What comes next.
In rough order of priority.
Wraith evaluation
Score the next detector on the same held-out tests and publish the results.
Broader evaluation
More model families and more public benchmarks, plus commercial moderation services as baselines.
Independent validation
Outside reproduction of the results, starting from the logs.
Deception and scheming proof-of-concept
Test whether internal activity can reveal a model that knowingly misleads. A research question, not a capability.
A deployable monitor
For self-hosted and open models.
Scaling the data and models
More recorded activity and larger models behind the detectors.
How results are reported.
- Only held-out numbers. No figure comes from prompts the detector trained on.
- Name the benchmark. Every number states its test set and the false-alarm bar it was measured at.
- Say what hasn’t been measured. Limits, untested models and missing comparisons sit next to the results.
Talk to Ulterior Safety.
Investors, AI labs, teams deploying open models, and press. Messages go straight to the founder.