The one-stop shop for LLM probes
See inside your model while it generates.
Runtime model monitoring with activation probes, starting with LLMs and video models. Catch risk during generation with our probes, or ones we train on your policy. For teams hosting their own open-weight or proprietary models.
Probe Library
Ready-made activation probes for unsafe content, code vulnerabilities and adult video content, each with published benchmarks and known limits.
Learn moreCustom Probes
Probes trained on your model and your own risk categories, with consulting from the researchers who build them.
Learn moreProbe-Fitted Models
The open-weight or proprietary model you choose, delivered with probes fitted to it.
Learn moreUndercurrent
Our open-source runtime for probes on LLMs during generation, on Hugging Face Transformers or vLLM (experimental).
Learn moreWho it's for: teams that host their own models. Activation probes read a model's internal state, so they won't work on a model you can only reach through a closed API.
How it works
The model already knows. We just read it.
Generation runs as normal.
The model processes the prompt and begins generating a response.
A probe reads the activation.
A small model (typically under 0.2% the size of the base model) inspects the internal state at inference time.
A risk score returns inline.
No separate guard model, no separate scanning pass — the signal is available before generation completes.
Peer-reviewed research
Safety Beyond the Interface: Detecting Harm via Latent States in Large Language Models
We kept running into the same frustration with external guardrails: they're blind. They see tokens — what went in, what came out — but nothing in between. So we asked a simpler question: does the model already know when a prompt is harmful? We trained lightweight probes on LLaMA-3.1-8B's hidden states and found that it does — competitive with 7B guard models at a fraction of the cost.
IEEE XploreLatent Space Probing for Adult Content Detection in Video Generative Models
Video generation models can be coaxed into producing explicit or harmful content with little effort. We trained lightweight probes on the internal latents of CogVideoX to detect unsafe content before a single pixel is decoded — competitive with 8B guard models at 500×+ lower latency.
IEEE Xplore (DOI)Probing the Prefill: Detecting Code Vulnerabilities via Latent Activations
AI-generated code ships faster than anyone can review it. We asked whether a model's own hidden states, read at the moment it writes code, already carry a signal about whether that code is vulnerable — and trained lightweight probes on those latent activations to find out.
arXivBuilt for where safety matters most
LLM Safety
Flags harmful prompts and responses from a chatbot or assistant's own activations, without a separate guard model.
Learn moreAI Coding Tools
Early research: a coding model's activations carry a signal about vulnerable code. 68.8% best-case F1 and 41.7% on average, on labeled code, not yet live completions.
Learn moreGenerative Video & Media
Probes on video diffusion models flag explicit content before a frame is decoded.
Learn moreSee it on your own model
Use our probe-fitted models, or keep yours. Talk to us about running a pilot on your stack.
Talk to us
