Investors
Raising early funding.
Ulterior Safety is raising early funding to scale our evaluations and extend detection to deception. Detecting deception is a research direction today; the funding is to find out how far the approach goes.
01Why now
Output checks are no longer enough.
-
Safety training can be removed
Open models can be modified so that they stop refusing harmful requests. Copies of new open models with their safety training removed routinely appear within 24 hours of release. Once that happens, the model’s own safety no longer protects anyone.
-
Output filters miss internal intent
Filters that read only the text cannot see what the model is doing internally.
-
A check that does not depend on the model’s safety
Ulterior Safety’s detector reads the model’s internal activity, so it keeps working when built-in safety has been stripped out.
02What exists today
A working detector, with held-out results.
Ghost P beats two leading open guard models on our test: the same 1,259 clearly harmful prompts, at the same bar of 1% false alarms.
95.4%
caught by Ghost P, on a model with safety training removed
89.2%
caught by Llama Guard 3 8B
46.6%
caught by ShieldGemma 9B
The test set is Ulterior Safety’s own and favours Ghost P somewhat; on two smaller overlap sets Llama Guard 3 was ahead. One model family tested, no external audit yet. Full results and limits.
- A full research pipeline built in-house: running models, recording their internal activity, and training and evaluating detectors.
- The next detector generation, Wraith, in development.
03What’s next
Where the funding goes.
-
Scale the evaluations
More model families, more public benchmarks, and independent validation of the results.
-
Extend detection to deception Research direction
A proof-of-concept on whether internal activity can reveal a model that knowingly misleads.
-
Towards a deployable monitor
For self-hosted and open models.
One-pager
Request the one-pager.
A short summary of the company, the results and the plan, sent by email.