Probe library

Activation probes that read safety signals from inside the model

Probes for unsafe content and code vulnerabilities in LLMs and adult content in video models, each backed by peer-reviewed research.

3
Probe families across LLMs, coding models and video models
<0.2%
Typical probe size relative to the base model
<10ms
Overhead: under 1ms on LLMs, 4–6ms per 10s video clip

Small

Roughly 11–16M parameters, not another multi-billion-parameter model

Inline

Reads activations the model already computes during generation

Published

Benchmarks and limits are public, in papers you can check

Extensible

Trained for any open-weight or proprietary model

Reads internal state

External filters see only tokens or pixels. Probes read what the model computes on the way to its output.

How to get them

Library probes are delivered fitted to the open-weight or proprietary model you host, through a pilot or engagement. There is no public download yet.

Measured on your model

Results vary by model and dataset, so we evaluate probes on the model you plan to run.

Need a different policy?

Our enterprise offering trains custom probes on your own risk categories.

The probe library

IEEE DSN-W 2026

Unsafe content

Detects harmful intent in prompts and responses from an LLM’s hidden states.

Validated on Llama 3.1 8B

  • 99.1% F1 on WildJailbreak
  • 83.5% F1 on AEGIS 2.0
  • 82.7% F1 on BeaverTails
  • 12.6M parameters

Trails the strongest 7B guard models on BeaverTails. WildJailbreak result is on the vanilla subset, which tests intent rather than adversarial jailbreak techniques.

Read the research
IEEE Milcom 2026

Code vulnerability

Reads a coding model’s activations for signs that code is vulnerable.

Validated on Coding LLMs, including Qwen3.5-9B

  • 68.8% F1 on Devign (best result; published SOTA 67.9%)
  • 44.6–59.5% F1 on Big-Vul
  • 13–16M parameters

Tested on labeled corpus code, not yet live completions. Average F1 across 16 model/dataset pairs is 41.7%.

Read the research
IEEE DSN-W 2026

Adult content (video)

Scores a video diffusion model’s latents before a frame is decoded.

Validated on CogVideoX

  • 97.29% F1
  • 4–6ms per 10s clip
  • 11.3M parameters

Trained on encoded real-world video; the gap to model-generated latents is not fully characterized.

Read the research

Put these probes on your model

Talk to us about the model you run and the risks you need to catch.

Talk to us