Llevar la fiabilidad autónoma a enclaves seguros

Security review starts with three questions: where does test data live, who can reach the execution environment, and what leaves the network. Multi-tenant SaaS that ingests production-like data answers all three wrong, even when the vendor is reputable and well-funded.

Autonomy raises the stakes. A passive tool stores data; an agent observes, decides, and acts inside your systems. Without boundary-aware design, autonomous reliability stops being an asset and becomes an unbounded liability your CISO has to underwrite. This is why the secure enclave is a deployment model, not a configuration flag.

Intelligence and orchestration run in a control plane your security team can assess. Test and remediation execution run inside your enclave, private cloud, or on-prem footprint, where data never crosses an unapproved boundary.

The split is deliberate. The model that plans work and the runtime that touches sensitive systems are different trust zones with different controls. The System Graph and the governance layer live outside; the workload stays inside.

Control plane (policy, graph, orchestration)
        | signed work packages only
        v
Customer enclave: Edge Runners + local evidence
        | sanitized egress
        v
Aggregated telemetry (no raw customer data)

Work sent to enclave runners arrives as signed capsules: scoped commands, timeouts, allowed endpoints, and data classification labels. Runners reject anything unsigned or out of policy, so the control plane cannot push work the enclave has not agreed to accept.

The capsule is the contract. It is also the audit record: every action an agent takes inside your perimeter traces back to a signed, scoped instruction rather than an opaque model decision.

Edge Runners execute capsules against internal URLs, desktop clients, and private APIs that never resolve outside your network. They stream artifacts to local evidence stores, not to arbitrary vendor buckets.

Runners are the only component with reach into sensitive systems, which makes them the right place to concentrate isolation, least-privilege identity, and network policy.

Customers define what may egress: pass/fail summaries, redacted traces, content hashes, or nothing at all. Transfer policy is enforced at the boundary and recorded, so what left and under which rule is always answerable.

Default-deny is the correct posture. The enclave should treat egress as an exception that policy explicitly permits, not a convenience that operations quietly enables.

Screenshots, HAR files, traces, and logs remain in customer-controlled storage by default. Reviewers reach evidence through the security tooling they already operate, not a vendor console outside the perimeter.

Evidence stays where the data does. That keeps retention, access, and discovery inside the controls your auditors have already approved.

When telemetry leaves the enclave, it is minimized and scrubbed before it crosses the boundary. The goal is operational visibility, fleet health, run status, aggregate pass rates, without exfiltrating sensitive payloads.

Sanitization is performed inside your perimeter, on the runner side, so the control plane never sees raw customer data even momentarily.

Runners integrate with privileged access management and secret vaults using short-lived credentials, with no long-lived keys held in vendor SaaS. Secrets are injected at execution time and never appear in agent prompts, capsule payloads, or external logs.

This matters because ~80% of developers bypass security policy under delivery pressure, per our analysis. Governed runners remove the path of least resistance: there is no convenient place to paste a credential where it can leak.

Concretely, a single run proceeds in stages, each crossing a boundary on purpose. The control plane plans the work from the System Graph and emits a signed capsule. An Edge Runner inside your enclave validates the signature, pulls short-lived credentials from your vault, and executes against internal endpoints. Artifacts land in your local evidence store; only a sanitized summary egresses.

For a deeper, end-to-end view of how a fleet plans, executes, and produces evidence, see Inside a Zof run. The enclave changes where each step happens, not what governed autonomy does.

Every claim above has to be answerable after the fact. A defensible enclave produces a contiguous record from capsule to evidence to egress, with no gap a reviewer has to take on trust.

A fair challenge from engineering leaders is that a hard execution boundary will starve the agents of context and slow everything down. In practice it does not, because the boundary separates intelligence from data, not intelligence from outcomes. The control plane plans from the System Graph and from sanitized telemetry; it does not need raw payloads to decide what to validate next.

The enclave does not make autonomy weaker. It makes autonomy something a regulated buyer can actually authorize.

The right boundary depends on your residency and sovereignty posture. The control plane stays assessable in every model; what moves is where execution and evidence live.

Ask for reference architectures, data-flow diagrams, and named failure modes, not marketing claims. Then validate runner isolation, capsule signing, egress policy, and evidence retention in your own environment before you trust any of it. The reliability patterns differ by sector, and the right baseline for a bank is not the right baseline for healthcare; reliability patterns by industry is a useful frame for that conversation.

Autonomous reliability can run inside secure enclaves when the architecture respects the separation of intelligence and execution. Regulated buyers, especially in financial services, should demand this by default and verify it in their own environment, not accept it as a bespoke project.

If the boundary is real, the agents propose and your people authorize, with every action signed, scoped, and recorded. That is the version of autonomy a security review can sign.