Why the CISO’s gut says ‘no’ until you see the ledger
The first thing a security lead notices is the absence of an immutable record of what an autonomous process has done. Without a tamper‑evident audit, every privileged call becomes a black box that can be replayed, altered or omitted after the fact. This missing provenance is a direct conflict with the EU AI Act’s requirement for traceability and with DORA’s insistence on auditable automation.
Imagine an agent that provisions a storage bucket for a machine‑learning job and then silently switches the bucket’s policy to public. If the action is never written to a reliable ledger, the breach can remain undiscovered until a regulator questions the data‑flow. The very act of not recording the decision defeats any reasonable defence in a compliance audit.
Ungoverned agents: the hidden attack surface
Customers can rent GPUs and VMs from a cloud provider, but the provider often supplies no product that makes those agents safe to run. An ungoverned agent can issue tool calls, invoke LLMs and execute code without any policy checks, opening a pathway for prompt‑injection attacks or for the agent to retrieve cryptographic keys from a shared secret store.
Suppose an agent receives a user‑generated prompt that contains a crafted instruction to retrieve the contents of a configuration file. Without a prompt‑injection guard, the agent forwards the request to the underlying model, which then leaks the file’s contents back to the attacker. The leakage is invisible unless a DLP filter or a guard is in place.
- No governed agent product for customer‑rented GPUsMAQPNA supplies default‑deny policies and a gateway that checks every tool and model call before it reaches the GPU
- Hard tenant isolation of sandboxes and keysEach agent runs in its own gVisor sandbox and receives a short‑lived Ed25519 token with a SPIFFE subject
- Metered usage per tenant for billing and complianceHierarchical budgets and token‑rate limits put a hard ceiling on spend while the ledger records every charged action
- Proof of data‑sovereignty for regional customersTool pinning and data‑residency rules, combined with an audit‑ready ledger, give concrete evidence of where processing occurred
- Rapid, safe shutdown of rogue agentsA kill switch revokes every agent everywhere in 302 ms at p99, providing an emergency stop that respects sovereign boundaries
Tenant isolation – more than a network slice
Hosting many customers’ agents on the same physical infrastructure demands hard isolation of sandboxes, keys and records. A single compromise in one sandbox should never allow a read of another tenant’s ledger or a sideways call to a private API.
A plausible failure: an agent in Tenant A discovers a mis‑configured container runtime and escapes to the host, then reads the SPIFFE token of Tenant B and uses it to call a billing service. The breach is possible when the isolation layer relies only on network policies rather than on a hypervisor‑level barrier.
Each agent session running in its own gVisor sandbox adds a lightweight yet strong boundary that prevents system‑call escape. Coupled with short‑lived Ed25519 identity tokens, the risk of token replay across tenants is dramatically reduced.
Metering, sovereign billing and the EU AI Act
A cloud operator must charge per tenant for sandbox time, governed tool calls and model tokens. Without hierarchical budgets and token‑rate limits, an errant agent can consume resources indefinitely, inflating the bill and exhausting capacity for other customers.
Imagine a scenario where an agent repeatedly invokes a high‑cost language model because its internal loop lacks a back‑off policy. The spending spikes, the operator cannot attribute the cost to a specific tenant, and the regulator questions whether the provider exercised “reasonable oversight”.
EU AI Act compliance also demands evidence that data never left the declared region. Tool pinning and data‑residency rules, together with a ledger that records the exact version of every tool used, provide the artefacts needed for a sovereign audit.
Putting the pieces together with MAQPNA
MAQPNA supplies the missing governance layer that the cloud environment itself does not provide. A single gateway sits between agents and every downstream resource, enforcing a default‑deny policy per tool and per argument. Only calls allowed by an explicit rule are permitted, and any risky call is held for a four‑eyes approval process.
The runtime writes every decision to a SHA‑256 hash‑chained, tamper‑evident ledger; a call that cannot be recorded is blocked. The ledger is the evidence pack that maps directly to the EU AI Act, DORA and other sovereign requirements, supporting an audit without claiming certification.
In practice, a CISO can enable the kill switch, which revokes every agent everywhere in a cross‑region test in 302 ms at p99, and rely on the built‑in DLP and prompt‑injection guards to stop private data from leaving the sandbox. The additional latency of policy checks—2.2 ms at p99 under load—remains well within typical cloud SLA tolerances.
- Default‑deny per tool / argument
- Four‑eyes approval for risky calls
- SHA‑256 hash‑chained ledger
- gVisor sandbox per session
- Hierarchical budgets and token limits
- Tool pinning and regional routing
Takeaways
- Configure a default‑deny policy and list only the tool calls each agent truly needs.
- Enable four‑eyes approval for any call that writes to storage or contacts an external LLM.
- Assign each tenant its own gVisor sandbox and short‑lived SPIFFE identity token.
- Set hierarchical budgets and token‑rate limits before any production rollout.
- Export the SHA‑256 ledger regularly to satisfy EU AI Act traceability checks.
Questions people ask
How can I stop an agent from leaking data across tenants?
Deploy each agent in a gVisor sandbox, enforce default‑deny policies, and use the built‑in DLP guard that blocks the lethal trifecta of private data, untrusted content and outbound channels.
What evidence does MAQPNA provide for an EU AI Act audit?
The runtime generates a hash‑chained ledger of every tool, model and agent‑to‑agent call; this ledger can be mapped to the Act’s traceability requirement via the supplied evidence packs.
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 data centres and clouds. Spotted a mistake? Tell us.