Attackers win because they control the input. CyberProfiler relies on something they can't control - internal model activations. We turn 1M activations per token into a full trace and classify it in under 150ms. (no input data retained)
We have identified a reliable 'malicious threat' signature in activation space which CyberProfiler classifies on. This gives defenders the upper-hand on prompt injection because it generalises well to novel and emerging threat patterns.
- Quickstart: get a key and analyse your first prompt with the Python SDK.
- Analyse a prompt: the API reference, where you can try it.
- Python SDK reference: every option in the Python SDK.
To get access, visit the Cyber Profiler page.