CAI Technology
Book a demo
Menu ☰
janus pilot

JANUS — AI Control Threshold · Real-time DLP for prompts

The lock, not the camera. The decision moves out of the audit and back into the flow.

Request a demo →
The problem

Visibility tells you what happened after the data has left. JANUS stops the leak in real time. The only moment you can decide is before the prompt leaves the organization — the rest is clean-up.

How it works

  1. 1

    Every request to any AI (Claude/Gemini/ChatGPT/Copilot/local model) passes through JANUS before leaving the network. Transparent mode — no client change needed.

  2. 2

    Detection combines deterministic scanning (national IDs, IBAN, API keys, secrets) with locally-run context-aware classifiers for unstructured text — commercial strategy, vulnerability descriptions, names mid-sentence.

  3. 3

    Mode 1 (Anonymize) — JANUS masks sensitive data with substitutes before the request leaves, then restores real values on the way back. Frictionless for the user.

  4. 4

    Mode 2 (Block) — three policies per data category — (a) block-and-explain (b) warn-and-let-decide (c) block-and-escalate to security/DPO.

  5. 5

    Inspection runs on the client's infrastructure, in the EU. Identifiable data never leaves the perimeter in clear text — neither to the external model nor to CAI Technology.

JANUS diagram — user pastes a prompt, JANUS anonymizes or blocks in real time, identifiable data never leaves the EU on-premise perimeter; the response returns in clear text to the user.
JANUS diagram — user pastes a prompt, JANUS anonymizes or blocks in real time, identifiable data never leaves the EU on-premise perimeter; the response returns in clear text to the user.

Capabilities

Threshold layer for any AI

Works with Claude, Gemini, ChatGPT, Copilot, or any local model — without changing the client. Transparent mode, deploy in a few hours.

Reversible anonymization

Sensitive data is masked with substitutes before the request leaves. On the response, real values are restored — the user feels no friction.

Three block policies per category

(a) block-and-explain, (b) warn-and-let-the-person-decide, (c) block-and-escalate to security/DPO. Applied selectively per data category.

Inspection runs on your infrastructure

The filter that protects your data does NOT take your data out of the perimeter to analyze it. Sovereignty by architecture, not by addendum.

Hybrid detection on unstructured text

Deterministic rules for national IDs/IBAN/API keys + locally-run context-aware classifiers for unstructured text (commercial strategy, vulnerabilities, names mid-sentence).

Audit trail per request

Every decision (anonymize/block/escalate) is logged with context. Reduces GDPR/NIS2/DORA compliance proof to a log search.

Tech stack

  • ▸ Rust core inspection engine (sub-millisecond per request)
  • ▸ PostgreSQL audit log (per-decision context)
  • ▸ Local LLM classifier (Qwen3 / gpt-oss family — context-aware on unstructured text)
  • ▸ Deterministic detectors (regex + finite-state) for CNP, IBAN, API keys, JWT, OAuth tokens
  • ▸ Reverse-proxy mode (transparent HTTP/SOCKS) — works behind Caddy/HAProxy/nginx
  • ▸ OIDC + RBAC via CAI-AUTH for admin / DPO console
  • ▸ Deploy: Docker Compose pe 1 VM până la HA cluster (Swarm sau k8s)

Evidence

  • ✓ Active in internal pilot on real traffic — catches, masks and stops exactly what it should.
  • ✓ Identifiable data does NOT leave the EU on-premise perimeter — neither to the external model nor to CAI Technology.
  • ✓ Decision moves out of audit and back into the flow — the moment of intervention is before the leak, not after.
  • ✓ Three policies per data category — the same company can hard-block an API secret and anonymize a customer's name, simultaneously.
  • ✓ Per-request audit trail, aligned with GDPR Art. 5, NIS2 Art. 21, DORA, EU AI Act Art. 13 and OWASP LLM02 2025.

Deliverables

  • ✓ 2-week discovery sprint — integration architecture + PoC on your real traffic (anonymize + block, both modes).
  • ✓ Policies per data category, written jointly with the DPO/CISO — not copy-paste from another client.
  • ✓ Deployment on your infrastructure (on-premise or EU private cloud), reproducible build.
  • ✓ Admin/DPO dashboard with full decision log (anonymize / block / escalate) and per-category statistics.
  • ✓ DPO runbook — how to respond to a transparency request, an incident, a NIS2 audit.
  • ✓ 90-day support window for fine-tuning detectors on your corpus.

Typical timeline

2-week discovery sprint · 4–6 weeks productive pilot (deploy + tuning) · GA within 2–3 months from kickoff.

FAQ

Why isn't a visibility tool enough? +
Visibility tells you what happened. JANUS stops things in real time. Data leaked into a prompt cannot be recalled after it reaches an external model — the only moment you can decide is before. One doesn't replace the other; you want both.
How does it integrate with existing AIs? +
JANUS sits on the network path between user/app and the external AI endpoint. For most cases (Claude, Gemini, ChatGPT, Copilot, etc.) it's transparent — no client change needed. Same for local models.
Does our data leave the perimeter to be analyzed? +
No. Inspection runs on your infrastructure (on-premise or EU private cloud). The filter that protects your data does not, itself, take your data out of the perimeter. Neither to the external model nor to CAI Technology.
What metrics (recall, FP rate, latency) do you guarantee? +
We're in pilot and measuring continuously. We'll publish fixed numbers once we exit pilot — not before. Honestly, detection on unstructured text is the active research zone. Bring a representative set and we'll measure on your data.
How does it align with GDPR, NIS2, DORA, AI Act? +
GDPR Art. 5 (data minimization at the source), NIS2 Art. 21 (risk management in the control), DORA (third-party ICT risk), EU AI Act Art. 13 (transparency in the decision log), OWASP LLM02 2025 (sensitive info disclosure). Architectural alignment by design, not certification. See our article under /articles/.
How do I start a pilot? +
Write us via /contact/ with subject 'JANUS pilot'. We reply within 24 working hours. Discovery sprint 2 weeks — architecture + PoC on your data — then joint decision.

We start with a 30-minute conversation.

Free AI-readiness audit for companies with 50+ employees. We reply within 24 hours.

Open contact form →