Data‑sovereign execution: why the agent must stay in your VPC
Regulated buyers refuse to ship raw training data to a SaaS provider; they demand the reasoning engine live inside their own virtual private cloud or on‑prem data centre. The EU AI Act Art. 12 makes data‑residency a legal requirement, and a breach of that rule can invalidate an entire model deployment. That reality turns a convenient "host‑the‑agent‑elsewhere" strategy into a compliance nightmare.
The simplest control is to place the agent runtime inside the customer's own cloud account, Kubernetes cluster or data centre. MAQPNA provides a gateway that sits between the agent and every external tool, model or service, ensuring that no call leaves the sovereign perimeter without explicit approval.
Imagine an agent that reads support tickets to prioritise a bug‑fix queue, but the SaaS vendor streams the raw tickets to a third‑party analytics bucket. The bucket resides outside the customer's jurisdiction, violating Art. 12 and exposing private customer data to an unauthorised party.
- Deploy MAQPNA inside the VPC that hosts the production workload.
- Disable any outbound internet egress that is not covered by a signed policy.
- Audit all DNS resolutions from the agent runtime for unexpected domains.
Per‑call least‑privilege: policing tooling and arguments
Security teams ask the familiar question – "Which tools can the agent call, with which arguments, and who can stop it?". In an AI‑centric organisation, agents often need to invoke internal services such as a payment processor, a model‑registry, or a CI pipeline. Granting blanket access to these endpoints creates a wide attack surface that is difficult to review.
MAQPNA implements a default‑deny policy per tool and per argument. A call is allowed only when a rule explicitly permits it, and any call that does not match a rule is blocked. High‑risk calls are held for a person to approve, with four‑eyes approval and separation of duties, ensuring that no single actor can silently trigger a destructive operation.
Suppose an agent is permitted to invoke the endpoint "/delete‑user" with any user‑id argument. An attacker crafts a prompt that includes the identifier of a senior admin account; the agent dutifully forwards the request, and the privileged account disappears without any human oversight.
- Create a whitelist of allowed tool‑names (e.g. payments, model‑lookup).
- Define argument schemas that reject unexpected formats or out‑of‑range values.
- Enable four‑eyes approval for any call that mutates critical state.
- Run agents inside the customer’s own cloud with no external control planeMAQPNA runs in the customer's own cloud account, Kubernetes cluster or data centre, providing a sovereign runtime.
- Enforce per‑call least‑privilege with explicit allow listsDefault‑deny policy per tool and per argument, with four‑eyes approval for risky calls.
- Guard agents against prompt‑injection from untrusted text sourcesPrompt‑guard screening can be enabled to filter inputs before they reach the LLM.
- Provide an immutable, tamper‑evident record of every agent actionEvery decision is written to a SHA‑256 hash‑chained ledger; a closed audit‑failure policy blocks unrecorded calls.
- Terminate all agents instantly in an emergencyA kill switch revokes every agent, everywhere; in a cross‑region test it took effect in 302 ms at p99.
- Prevent silent tool version changes that could undermine policiesTool pinning catches a "rug pull" by enforcing pinned MCP tool definitions.
Prompt‑injection as an attack surface for read‑only agents
Even agents that are only meant to read emails, tickets or web pages can be steered by the content they ingest. An attacker who can influence those sources can embed a hidden command that the LLM interprets as an instruction, a classic prompt‑injection scenario. The risk is amplified when the downstream tool call can exfiltrate data or trigger a workflow.
MAQPNA offers built‑in Prompt‑Guard screening that can filter inputs before they reach the LLM. The guard can be configured with known patterns, black‑lists and sanitisation rules, reducing the chance that malicious payloads slip through. It works together with the default‑deny gateway, so a malicious prompt that tries to invoke a tool will be blocked unless a policy explicitly allows it.
A plausible failure: an engineer forwards an internal Slack message that contains the phrase "export all logs to s3://public-bucket". The agent parses the message, the Prompt‑Guard does not recognise the phrase as malicious, and the agent subsequently calls the storage API, leaking confidential logs.
- Enable Prompt‑Guard on every inbound text source.
- Regularly audit the guard rule‑set for false‑negatives.
- Combine guard output with the default‑deny policy to enforce a second line of defence.
Immutable audit trail versus mutable logs
Auditors demand a record of every automated action, not a screenshot or a log file that can be edited after the fact. Traditional application logs are mutable and often stored on the same filesystem that the compromised process writes to, making them unreliable as evidence.
MAQPNA writes every decision to a SHA‑256 hash‑chained, tamper‑evident ledger. The audit‑failure policy can be set to "closed", meaning that if a call cannot be recorded it simply does not run. In a benchmark the system produced 868,473 records without dropping a single entry, proving that the ledger can keep pace with high‑throughput workloads.
Imagine an attacker who disables the central syslog daemon to hide their tracks. Because the agent’s call never reaches the ledger without a successful write, the policy blocks the operation, and the audit log shows exactly why the request was rejected.
- Configure the ledger with the "closed" failure policy.
- Export ledger snapshots to an immutable storage tier nightly.
- Cross‑check ledger entries against external compliance frameworks (EU AI Act, DORA).
Putting the controls together: a Monday‑morning checklist
When the week starts, the CISO should verify that the agent runtime is still operating inside the authorised VPC, that the kill‑switch is reachable, and that no new tool definitions have appeared without a pin. Running the MAQPNA preflight command against the production profile surfaces any drift before traffic begins.
Next, review the default‑deny policy file: confirm that only the minimal set of tools and argument patterns are listed. Check the pending approvals queue – any lingering high‑risk calls indicate a process bottleneck that needs remediation. Finally, scan the latest ledger export for anomalies such as a sudden burst of calls to a newly‑added endpoint.
A final sanity check is to fire a test request that attempts a prohibited action (for example, a delete‑user call with a random ID). The response should be an explicit deny, and the ledger must contain the denial record.
- Run
maqpna preflight --profile prod -f values.yaml --stricteach Monday. - Audit the pending‑approval list for any stale high‑risk entries.
Takeaways
- Deploy MAQPNA inside the VPC that hosts production AI workloads.
- Adopt a default‑deny policy and require four‑eyes approval for any mutable call.
- Enable Prompt‑Guard on every inbound text source the agent processes.
- Set the ledger to "closed" failure mode and verify daily export integrity.
Questions people ask
Can I let an AI agent call external APIs from my VPC?
Only if a policy explicitly permits the tool name and argument pattern. Unlisted calls are denied by the default‑deny gateway, keeping the outbound surface under control.
What evidence do auditors need to see that an agent acted safely?
The SHA‑256 hash‑chained ledger provides an immutable, time‑ordered record of every decision. Exporting the ledger to immutable storage satisfies most AI‑Act and DORA audit requirements.
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 ai companies. Spotted a mistake? Tell us.