
News
OpenAI Publishes a Supervised Lab Design for Daybreak Red
OpenAI's November 19 guide isolates a supervised Daybreak Red investigation: local Codex, a separate agent VM, network limits, and logs.
OpenAI has published a reference architecture for supervised security investigations with Daybreak Red. The document is about the lab around the model: who directs the work, where commands run, and which records sit outside the agent's reach.
The learn page, dated November 19, 2026, is titled "Deploying Daybreak Red" and calls the piece "a reference architecture for supervised security investigations." OpenAI filed it under Guides and linked a PDF covering architecture, safety controls, environment configuration, and how to run and monitor an investigation. The PDF header is dated November 9, 2026. OpenAI is the listed author.
The PDF describes Daybreak Red models as "powerful cybersecurity investigation tools" and treats the architecture as a starting point. "The most important output is understanding the controls so you are able to safely implement OpenAI cyber models in your environment."
A workstation that directs, a VM that executes
Codex on the analyst's workstation directs the investigation and reviews results. Commands run on a separate agent VM. Lab network rules limit what that VM can reach. A separate log store records what Codex did and which requests reached the test app.
The PDF names eight components:
- Analyst. The authorized person who starts the investigation, configures and scopes the test, stops the run, and analyzes results.
- Isolated investigation environment. A logical separation. The reference design describes hosted VMs in an isolated VPC. "The specifics of this environment will depend on your infrastructure."
- Coordinator. A local Codex install on the analyst's machine. It manages the connection to the remote host, presents results, and supervises the run. The analyst reaches the remote Codex installation on the agent VM through this instance.
- Agent VM. Runs Codex as a non-admin user inside the isolated environment and executes the analysis. The guide uses a disposable VM. A container is acceptable only if it enforces the same file and network restrictions.
- Model inference. Connectivity to OpenAI for model serving. The PDF also notes that OpenAI models are available through AWS.
- Network controls. A configuration point for resource access, via iptables on the agent VM, a firewall, or security groups.
- Target. The non-production application under investigation.
- Telemetry. A data store outside the target and the agent VM that persists logs of what happened and when.
The architecture diagram also labels SSH, an action gate called a guardian, and three record streams: execution logs, network and identity telemetry, and app and access logs. The fetched prose does not define the guardian label further.
Safety controls
The design follows what the PDF calls OpenAI's recommended configuration for cyber work. The stated goal is enough access for the task, with visibility and governance over model activity.
Scope defines the target, permitted actions, investigation window, and stopping point. It also covers what evidence is required, and whether Codex uses anonymous access or a dedicated account. If Codex signs in, that account is limited to the role and resources the investigation calls for.
On the VM, Codex runs under a dedicated non-admin account. The Codex sandbox limits what commands can read and write. A writable workspace is for scripts and results; a read-only workspace can still allow approved network requests. Private keys, deployment credentials, and host-management sockets stay outside that account's reach. The build section says Codex uses bubblewrap for the file sandbox, and that the sandbox should be verified before an investigation starts. Higher-risk work should use a fresh VM and lab.
Network limits sit in more than one place. The Codex network proxy can narrow where sandboxed commands connect. In the PDF's example, the VM may be allowed to reach three staging apps while the Codex allowlist permits only the one in scope. Beyond that allowlist, separate controls the Codex account cannot change restrict access outside the investigation environment. The stated best practice is to allow only required destinations, and to check network access separately for any browsers, connectors, or MCP tools that are enabled.
When an action needs approval and is eligible for Auto-review, a separate reviewer decides whether it can proceed. The analyst sets the scope and supervises the run. If Codex needs a new target or account, or an action outside the agreed scope, the run pauses until the analyst confirms authorization.
Telemetry keeps Codex tool activity, target request logs, and application sign-in events in a log store outside the Codex account's control. That account cannot change or delete the stored records.
What the deployer still supplies
The build section assumes a non-production target prepared with synthetic data, a workstation with Codex, a lab VM, and a separate log store. If the investigation signs in, the guide calls for a dedicated test account, only the permissions the investigation needs, a short-lived credential, and reusable secrets kept out of Codex's reach.
The agent VM should reach only the approved test app and the services needed for Codex sign-in, model access, and log export. Sign-in uses a ChatGPT account and workspace approved for Daybreak Red, or, for organizations on API access, a key from the approved Daybreak Red project. An administrator places mandatory settings in requirements.toml so the VM account cannot override them. Codex logs leave the VM through an OpenTelemetry exporter; application request logs and sign-in events go to the same store.
Before an investigation, a preflight from the workstation is meant to confirm workspace write limits, the network allowlist, and that independent logs captured the checks. During a run, routine work inside the agreed scope can continue. The analyst pauses if Codex reaches a different host, encounters real user data, or asks for broader permissions, and stops the run if a required control or its monitoring fails. Before the first investigation, the guide says to verify that the agent VM can be stopped, or its access to the target blocked, from outside Codex. Ending a run means stopping requests and background jobs, confirming both have stopped, closing the remote session, and revoking temporary credentials.
An optional section applies the design to a disposable, non-production Gitea instance "to identify and exploit a known CVE." That walkthrough is not part of the architecture summary, and this article does not reproduce it.
The reference architecture does not stand in for a configured lab. It names the split — local Codex as coordinator, a separate agent VM, network limits, and an external log — and leaves the VPC, firewall, and log store to the team that deploys them.
Read the November 19 learn page and the linked PDF, and map those named controls onto an isolated environment before treating an investigation as in scope.
Sources
- Deploying Daybreak Red openai.com
- Supervised investigations with Daybreak Red: a reference architecture (PDF) cdn.openai.com