Platform

Zero-trust AI
in the network layer.

CareProxy sits in the network — not on every endpoint. Every request is classified, routed fail-closed, and cryptographically logged. 100% on the hospital's own infrastructure. No US cloud.


How zero-trust routing works

Three stages — Triage, Route, Verify. Every request is classified and routed in under 10 milliseconds.

01
Step

Triage

Classify in real time

Every request is scanned against our Clinical Domain Dictionary. Patient identifiers, ICD-10 codes, medication names, and department-specific terms are classified by risk level. Sub-10ms.

Status: Inspection
02
Step

Route

Zero-trust decision

High-risk payloads are locked to your on-prem AI — never public cloud. Low-risk payloads route to OpenAI/Anthropic. Fail-closed: when classification is uncertain, data stays local. Sticky-local sessions.

Status: Routing
03
Step

Verify

Cryptographic proof

Hash-chained audit log records per request: session ID, risk score, trigger rule, destination, SHA-256 payload hash. Immutable. Verifiable. Ready for auditor and regulator.

Status: Evidence

Stage 02b

AI-injection-firewall

Ten attack patterns · per-persona · ALLOW / WARN / BLOCK

Patient notes and clinical documents pasted into a prompt may contain hidden instructions — deliberate or accidental — attempting to manipulate the model's behaviour. CareProxy scans every prompt for ten attack patterns before it ever reaches an external model.

Three-tier decision ALLOW WARN BLOCK

The ten signal types

  • 01 Instruction Override

    Attempts to cancel or replace the system prompt. E.g. "ignore all previous instructions and …".

  • 02 Role-Tag Injection

    Embedded chat-format tags forging a new system or assistant turn. E.g. ChatML <|im_start|>, Llama [INST], Anthropic-style turn prefixes.

  • 03 Exfiltration Request

    Attempts to make the model dump records, list everything, or reveal the system prompt. E.g. "list all patients", "repeat your initial instructions".

  • 04 Tool-Use Hijack

    Forged tool-call markup inserted into the prompt to trick the model into executing functions. E.g. fabricated function-call JSON or tool-use blocks.

  • 05 Unicode Tag Smuggling

    Invisible Unicode tag characters (Plane 14, U+E0000–U+E007F) and bidi-override marks that hide instructions inside otherwise plain text.

  • 06 Encoded Payload

    Base64-style blocks whose decoded contents read as instructions.

  • 07 Delimiter Breakout

    Markdown fences or triple-quotes used to break out of a quoted block and inject a new conversation turn. E.g. ``` followed by User:.

  • 08 Indirect Injection

    Instruction-shaped content embedded inside a quoted patient note, document, or context block — the classic supply-chain prompt attack.

  • 09 Multilingual Jailbreak

    Override phrases in non-English languages (Danish, German, Spanish, French, Chinese, …) used to slip past English-only filters.

  • 10 Length / Repetition Anomaly

    Single lines over 4 KB, or short tokens repeated to push the system prompt out of the context window.

Scoring is per-persona — clinical researchers get tighter thresholds than developer tooling that legitimately works with code. A Danish-language allowlist ensures clinical phrasings ("ignore the previous blood test, that was a measurement error") aren't falsely blocked.


Zero-Trust · On-Premise · Hash-Chained

Fail-Closed Architecture

CareProxy deploys 100% on-premise. Our Triage engine classifies and routes every request with zero data persistence. High-risk payloads are locked to your own AI models. Every routing decision is logged cryptographically — legal-grade evidence for CISOs, auditors, and regulators.

On-Premise Security Boundary
Classified
Hash-verified

On-Premise Deployment

Deploys inside your hospital's own ISO 27001-certified datacenter. Docker/K8s. No cloud dependency. Full data sovereignty from the first call.

Sticky-Local Sessions

Once a session is classified as high-risk, it is locked to local AI models — including subsequent calls from the same session. No leakage via context.

Fail-Closed Default

On classification uncertainty, routing error, or model timeout, data stays local. Zero-trust is the default — never the exception.

Hash-Chained Audit Log

SHA-256 hash chain per request, Ed25519-signed, persisted to Postgres. Both prompt and response are hashed. One-click forensic export. Optional RFC 3161 timestamp anchoring.

Architecture & Security

Technical specifications

Triage Latency < 0ms
Routing Zero-trust (fail-closed)
Audit Hash-chained SHA-0
Data at-rest 0% (stateless)
Deployment On-Premise (Docker / K0s)
Integration OpenAI-compatible REST API
Compliance GDPR, EU AI Act, ISO 0
Roadmap TEE Attestation (NVIDIA H0)

Hash-Chained Audit Log

Three layers of verification — from database to external authority

The log isn't a JSON file in a bucket. Every row is hash-chained, Ed25519-signed, persisted to Postgres, and can be exported as a self-contained evidence bundle that an external investigator can verify without contacting CareProxy.

01

What is logged?

  • Timestamp, session ID, decision, trigger rule, latency
  • SHA-256 hash of the prompt (payload_hash)
  • SHA-256 hash of the model's response (response_hash) — proves what came back, not just what went in
  • AI-injection-firewall score and signals (if any)
02

How is the log protected?

  • Each row is hashed together with the previous row's hash → hash chain
  • Each row is signed with Ed25519 (private key lives locally, never in the cloud)
  • The chain is persisted to Postgres — survives restart, can be audited independently from the database alone
  • Optional RFC 3161 timestamp anchor every hour at an external TSA — proves the log existed before a given point in time, independently of your own signing key
03

How is evidence handed over?

  • One click in the CISO dashboard produces a self-contained JSON file with the full chain, signature, public key, and verification instructions
  • An external investigator verifies with stdlib crypto (openssl, Python, Go) — no contact with CareProxy required

RFC 3161 timestamp anchoring is opt-in (TSA_URL env var). Most hospital networks block outbound HTTPS — turn it on once your network allows it.


Roadmap · Vision 2027

From routing to attestation — the path to Confidential Computing

CareProxy is built in two stages. Today: zero-trust routing between cloud AI and your own on-prem models — built for hospital CAPEX reality, not per-hour cloud rent. Tomorrow: when frontier providers (OpenAI, Anthropic) finally expose hardware attestation, CareProxy automatically becomes the verifier for public cloud AI. The moment hospitals can use the world's smartest models with cryptographic proof that data stays inside a hardware-sealed enclave.

Patent draft in progress
TEE H100
  1. Shipped 2026 · Stage 1 · Zero-Trust Routing

    On-prem classification and fail-closed routing: low-risk to cloud AI, high-risk locked to your own local LLMs (e.g. Llama 3). AI-injection-firewall in front of every model. Hash-chained, Ed25519-signed audit log persisted to Postgres — both prompt and response hashed. One-click forensic export. RFC 3161 timestamp anchoring (optional). Deployed with pilot partners.

  2. 2027 – 2029 · Stage 2 · Frontier Model Attestation

    Once OpenAI, Anthropic, and other frontier providers expose hardware attestation (TEE) via their public APIs, CareProxy automatically becomes the verifier for public cloud AI. Hardware-documented access to the world's smartest models — with the legal evidence in place.


Ready to take control of your AI infrastructure?

CareProxy is under active development. Join the waitlist to be notified as soon as pilot installations open.

Or reach out directly kontakt@careproxy.dk

Join the waitlist

Be among the first hospitals to get access to CareProxy's zero-trust AI routing. We'll reach out as soon as pilot installations open.