Ulterior Safety

Company

An independent AI safety research company.

Founder-led, self-funded, and focused on one question: what is the model doing inside?

01Mission

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; 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: whether the same approach can catch a model that knowingly misleads, hides what it is doing, or underperforms on purpose. That is a research direction at an early stage, not a capability.

02Founder

Jasper Jensen

Founder

Built
The full research pipeline
Funding
Independent and self-funded

Jasper Jensen founded Ulterior Safety and built its full research pipeline.

That work covers running models, recording their internal activity at scale, and training and evaluating the detectors, from the first prototype to Ghost P.

The company is independent and self-funded, and is raising early funding.

“Output filters miss what the model is doing internally. Ulterior Safety looks inside.”

The Ulterior Safety mission

03Roadmap

What comes next.

In rough order of priority. No dates are promised.

  1. Next detector evaluation

    Score the next edition on the same held-out tests and publish the results.

  2. Broader evaluation

    More model families and more public benchmarks.

  3. Independent validation

    Outside reproduction of the results, starting from the logs.

  4. Deception and scheming proof-of-concept Research only

    Test whether internal activity can reveal a model that knowingly misleads. A research question, not a capability.

  5. A deployable monitor

    For self-hosted and open models.

  6. Scaling the data and models

    More recorded activity and larger models behind the detectors.

04Principles

How results are reported.

  1. Only numbers measured on held-out data

    No figure comes from prompts the detector trained on.

  2. Name the benchmark

    Every number states the test set and the false-alarm bar it was measured at.

  3. Say what hasn’t been measured

    Limits, untested models and missing comparisons are stated next to the results.

Interested in the company?

Ulterior Safety is raising early funding.

For investors Get in touch