An antivirus for the age of AI.
PRAETOR LAYER is a runtime security layer for LLM apps and autonomous agents — WAF · EDR · DLP for the agent stack. It inspects prompts, data, model output and every agent action inline, and blocks what shouldn't get through.
Antivirus was built for files.
Attacks now travel in meaning.
A classic AV matches deterministic bytes. LLM attacks are semantic, probabilistic and contextual — the same sentence is safe or catastrophic depending on the agent's privileges. You can't sign a payload that rephrases itself infinitely.
So we keep what carries over — layered defense, hybrid detection, a live threat feed, quarantine, telemetry — and rebuild the rest as runtime application security.
Classic AV
- file signatures
- known-bad bytes
- static scan
- on-disk quarantine
PRAETOR
- semantic detection
- intent & provenance
- inline runtime
- action-level firewall
The threat model
Eleven classes across input, data, output, and agent actions — mapped to OWASP LLM Top 10 & MITRE ATLAS.
Cheap filters first. Expensive models last.
A hybrid conveyor — each stage only escalates what the previous couldn't resolve.
Normalize
base64 · hex · zero-width · NFKC · homoglyph
Signatures
regex + fuzzy TI feed
Embeddings
similarity to attack corpus
Guard model
fine-tuned classifier
LLM-as-judge
escalation for the grey zone
Try the working console.
The whole detection pipeline runs client-side in your browser. No data leaves the page.
Launch App ▸