Probe-fitted models

Your choice of model, with safety probes built in

Open-weight and proprietary models fitted with activation probes, so risk signals come from the model itself during generation.

83–99% F1
WildJailbreak 99.1%, AEGIS 2.0 83.5%, BeaverTails 82.7% (Llama 3.1 8B)
<0.2%
Typical probe size relative to the base model
<1ms
Probe overhead on LLMs vs. 50–500ms generation time

Any model

Open-weight or proprietary, fitted with probes for your use case

No separate guard model

The signal comes from activations the model already computes

During generation

Risk scores arrive before the response is complete

Three probe families

Unsafe content and code vulnerabilities for LLMs, plus adult content for video models

Bring any open-weight or proprietary model

Start from the model you already use or plan to use. We train probes on its activations and fit them to it.

Probes matched to your risk

Choose from our probe library, or train probes against your own policy categories.

Signal at inference time

A probe reads internal state as the model generates and returns a risk score inline.

Runs on your LLM serving stack

Undercurrent, our open-source runtime, attaches probes to LLMs on Hugging Face Transformers or vLLM (experimental).

Published results cover the specific models in our papers. Performance on other models is measured when probes are trained for them.

Tell us which model you want to deploy

We will scope the probes and evaluation for your model and use case.

Request a model