Section 9 of 25 2 min read

Execution Isolation and Network Egress

Bounding what a compromised agent runtime can reach, and enumerating every outbound path.

Objective

Bound what a compromised agent process can reach at the infrastructure layer.

Minimum bar (L2). Agents execute in isolated, non-privileged runtimes with no host filesystem and no host network access.

9.1 Controls #

IDControlMinimum bar (L2)Enforced state (L3)
ISO-1Enumerated egress pathsA documented list of permitted outbound destinations for every agent runtime, default-deny for everything else, with an explicit allowlist as the only route in. Each entry names the business reason it exists and the owner of the software terminating that path.Enumerated, owned, and monitored egress paths. Every outbound route—package proxies, artifact caches, telemetry endpoints, DNS resolvers, model API endpoints—is inventoried with a named owner and a patch cadence matching production applications.
ISO-2Egress segmented by trust tierAgents that process untrusted input—external retrieval, inbound email, web browsing, customer-supplied documents—run in a network segment distinct from agents holding privileged credentials. The separation is written into network policy and verified in the running environment at deployment.Egress domains segmented by trust tier. Agents processing untrusted input do not share an outbound path with agents holding privileged credentials.
ISO-3Ephemeral runtimesAgent execution contexts are rebuilt from a known-good image on a defined cadence rather than persisting indefinitely, and no agent writes to a shared filesystem that survives its task. State that must persist is externalized to a governed store with its own access controls.Ephemeral runtimes destroyed after task completion; no persistence between tasks unless explicitly authorized.
ISO-4Internal package mirrorPackage and dependency installation flows through an internal mirror or proxy rather than directly to public registries. That mirror is inventoried as a production asset with a named owner and is patched on the production cadence, not a best-effort one.Package and dependency installation from an internal mirror that is itself treated as a production security boundary, subject to vulnerability management, monitoring, and—critically—the assumption that it is a target.
ISO-5Egress anomaly monitoringEgress from agent runtimes is logged with destination, volume, and timestamp, and reviewed against a documented baseline on a defined schedule. Alerting exists for previously unseen destinations even where automated blocking does not.Egress volume and destination-diversity monitoring, with automated blocking on anomaly.

9.2 Evidence to Request #

  • The egress inventory with owners and last-patched dates.
  • Network policy for a representative agent runtime.
  • Evidence that the package proxy or artifact cache is in the same vulnerability management program as production applications.

9.3 Failure Modes #

  • One permitted egress path, running third-party infrastructure software, unreviewed because it is “just a cache.” This is the precise architecture that failed in July 2026: a single package-registry cache proxy carrying an undiscovered vulnerability, which frontier models found and exploited to reach the open internet from an environment believed to be isolated.4
  • Isolation asserted in design documents and not verified in the running environment.
  • Development, evaluation, and red-team environments given weaker isolation than production, despite frequently running models with safety controls reduced.

References cited in this section

1 of 32 · numbering matches the PDF

  1. 4"OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation," OpenAI, July 21, 2026.openai.com/index/hugging-face-model-evaluation-security-incident ↗
PDF