Ulterior Safety
See what AI models are hiding.
We read what AI models are doing inside, to catch harmful and deceptive behaviour before it reaches the output.
97% of harmful requests blocked on JailbreakBench, even on a model with its safety training removed.
01The problem
Checking what an AI says isn't enough
Anyone can strip the safety training out of an open AI model in minutes. Thousands of these modified models are already public. Filters that only read the output can't tell what the model intends.
02Approach
What we do
-
Look inside
We monitor a model's internal activity as it answers.
-
Catch it early
We flag harmful requests before the reply is finished.
-
Works when safety is stripped
Our detector still catches harm on models whose built-in safety has been removed.
Why Ulterior Safety Lightweight, runs alongside open models, and keeps working when safety training is stripped out.
03Evaluation
Early results
Our detector, running on a model with its safety training removed, blocked more harmful requests than the original model's built-in safety.
| Benchmark | Setup | Harmful requests blocked | Safe questions wrongly refused |
|---|---|---|---|
| JailbreakBench | Original model | 95% | 26% |
| Safety removed + our detector | 97% (better result) | 18% (better result) | |
| XSTest | Original model | 83% | 6% (better result) |
| Safety removed + our detector | 90% (better result) | 12% |
Early results on Qwen3.5-9B. Small benchmarks; we're expanding the evaluation.
04What's next
Next: catching deception
We're extending the same approach to scheming, where a model knowingly misleads, hides what it's doing, or underperforms on purpose.
05About
About
Ulterior Safety is a solo-founder company. I'm Jasper Jensen. I built the full pipeline myself: running models, recording their internal activity across 19,000 replies, labelling deception, and training detectors that work even when a model's safety has been removed.
AI is advancing fast, and models can already be pushed into helping with things they shouldn't, or into misleading the people who use them. Checking what a model says isn't enough; we need to see what it's doing inside. Ulterior Safety is independent and self-funded.
06Funding
We're raising early funding
To scale our evaluations and extend detection to deception.
07Contact
Work with us
Researchers, funders and teams deploying AI, we'd like to hear from you.
Message sent
Thanks, your message has been sent. We'll be in touch.