Heuristics
Deterministic regex & validators catch the obvious, private keys, AWS credentials, Luhn-valid cards, JWTs. Match or no match, no ambiguity.
CERBERUS
Cerberus inspects every outbound payload across four detection layers and returns a single verdict, allow, warn, or block, in under 100 milliseconds.
Every request flows through a fast-fail pipeline. Cheap deterministic checks run first; the heavy model only wakes when it has to. Each layer adds its own evidence, and Cerberus fuses them into one decision.
Deterministic regex & validators catch the obvious, private keys, AWS credentials, Luhn-valid cards, JWTs. Match or no match, no ambiguity.
Hybrid spaCy + exact-match recognition surfaces PII and company-specific entities, customer names, emails, phone numbers, national IDs.
FAISS-indexed embeddings of your proprietary source. Catches leaked code even after every variable has been renamed.
A DistilBERT model, ONNX-optimised for CPU, reads intent and sorts each payload into safe, suspicious, or exfiltration.
Scroll to watch a prompt pass through every detection layer, from keystroke to verdict.
Latest run against the expanded dataset of
labelled prompts, safe developer chatter,
accidental leaks, and outright malicious exfiltration.
Detections that landed correctly, per layer.
50 samples · classes fused to leak / no-leak.
Zero false negatives, nothing sensitive slipped through. The handful of false positives are deliberately conservative: Cerberus warns first.
Wherever a single pasted secret is a breach, Cerberus sits inline at the gateway. Here is how teams across different fields put it to work.
Most sensitive data never leaves through a breach, it leaves through a text
box. A pasted .env, a copied customer table, a screenshot of a
contract. Cerberus watches that text box.
Run the detection API, point your proxy at it, and watch the verdicts stream in.
Request a demo