Data residency can be broken by a single stray API call
When an AI‑driven help desk agent automatically forwarded a request to an external language model, the model processed the request in a jurisdiction that does not recognise the EU GDPR. The incident showed that an agent, even when embedded in a trusted front‑end, can become the conduit for cross‑border data movement without a human ever noticing.
Under EU AI Act Article 12 a high‑risk system must ensure that personal data never leaves the territory unless a specific safeguard is applied. The same principle is echoed by GDPR’s restriction on transfers unless adequate protections exist. Without a runtime that can enforce data‑residency policies at the point of tool invocation, the organisation risks regulatory breach and loss of public trust.
Audit trails must bind every decision to a human authorisation
EU AI Act Article 14 demands a clear record of who triggered an automated decision, which rule was applied and which human gave final sign‑off. In practice this means that an agent that classifies a benefit claim must leave an immutable entry linking the citizen file, the policy version and the approving officer.
A plausible failure: an agent automatically escalates a security incident, but the ledger entry is lost because the call bypassed the central gateway. The downstream analyst cannot prove whether the escalation complied with the internal SOP, exposing the agency to accountability gaps.
- Data must never leave the sovereign jurisdictionMAQPNA runs in the customer’s own cloud or data centre, so no external control plane can move data abroad.
- Every automated decision needs a tamper‑evident audit recordMAQPNA writes each call to a SHA‑256 hash‑chained ledger and blocks any action that cannot be recorded.
- Air‑gapped networks cannot rely on internet‑only servicesMAQPNA is designed for on‑prem installation, including an air‑gapped bundle that can be signed and loaded offline.
- Supply‑chain provenance of tools and models must be provableMAQPNA pins MCP tool definitions and alerts on any silent change, ensuring only verified binaries run.
- Spend and model usage must be capped and auditableHierarchical budgets and token‑rate limits in MAQPNA place a hard ceiling on consumption.
Air‑gapped environments need on‑prem runtimes, not cloud‑only agents
Classified networks in ministries are deliberately isolated from the internet. Conventional SaaS‑based agents that rely on external model endpoints cannot operate there, leaving a gap in automation that teams try to fill with ad‑hoc scripts that lack auditability.
Supply‑chain assurance for these networks requires knowing exactly which binary is executed, who built it and that no hidden update can replace it. Tool pinning and DLP that recognises the lethal trifecta of private data, untrusted content and outbound channels become essential guardrails.
Concrete Monday‑morning controls for a CISO
Start by inventorying every tool and model an agent may call. Define a default‑deny policy for each, permitting only those explicitly listed. Enable four‑eyes approval for any call that touches citizen data, ensuring separation of duties.
Activate the built‑in DLP to block the lethal trifecta and pin all tool definitions so that a silent version change raises an alert. Generate short‑lived Ed25519 identity tokens for each agent session and enforce hierarchical spend limits to stop runaway usage.
- Run a policy validation with
maqpna validate -f manifests/ -f policies/ --strictbefore any production rollout. - Schedule a daily review of the SHA‑256 ledger to confirm that no call has been omitted.
- Test the kill‑switch in a staging environment to verify the 302 ms cross‑region shutdown time.
Takeaways
- Define a default‑deny policy for every tool before any agent is allowed to run.
- Require four‑eyes approval for any call that accesses personal or classified data.
- Pin tool versions and enable DLP to stop the lethal trifecta of data loss.
- Run
maqpna preflight --profile prod -f values.yaml --strictas part of the CI pipeline. - Review the SHA‑256 ledger each day to confirm no unaudited action occurred.
Questions people ask
Can I use a cloud‑hosted LLM with MAQPNA for a classified workload?
MAQPNA can enforce data‑residency and DLP, but a classified environment that lacks internet connectivity requires an on‑prem model. The runtime itself can be installed offline to meet that need.
How does MAQPNA help satisfy EU AI Act Article 14?
Every agent call is logged to a SHA‑256 chained ledger with the identity token of the invoking agent and the approving human, providing the traceability the Act requires.
How this was written
Part of the MAQPNA sector series, written with Cloudflare Workers AI and checked automatically: every claim about MAQPNA comes from our fact sheet, no statistics are invented, and rules are named, never "certified". MAQPNA helps you produce the controls and evidence they ask for; your assessor decides the rest. See MAQPNA for government and public sector. Spotted a mistake? Tell us.