Quick Answer
OpenAI announced Daybreak Blue and Daybreak Red availability through Amazon Bedrock on August 11, 2026. This is an approved-access path, not public self-service access. Eligible defenders first need Daybreak Access approval; the exact AWS account must also have the selected model in a supported Region and valid Bedrock authentication.
Use Daybreak Blue first for most authorized defensive work. Daybreak Red is a separately approved specialist path for controlled vulnerability reproduction, exploit validation, penetration testing, red teaming, and complex system analysis. Approval for Blue does not grant Red.
What the AWS Launch Changes
The launch gives approved organizations a way to use Daybreak within existing AWS security, governance, procurement, and operating workflows. OpenAI says approved customers can access the model from the Amazon Bedrock console or through the Responses API using the bedrock-mantle endpoint.
It does not remove the need to verify:
- the approved OpenAI identity, organization or project, model, and product surface;
- the AWS account, IAM principal, Region, model access, quotas, and billing owner;
- the systems and actions covered by the security engagement;
- data classification, retention, logging, and escalation requirements;
- human approval before sensitive or irreversible actions.
Blue and Red at a Glance
| Decision | Daybreak Blue | Daybreak Red |
|---|---|---|
| Default role | Most authorized defensive workflows | Separately approved specialist cyber work |
| Example work | Triage, secure review, detection engineering, incident response, malware analysis in a controlled environment, patch validation | Controlled vulnerability reproduction, proof-of-concept validation, penetration testing, red teaming, complex system analysis |
| Approval | Trusted Access for the exact offering and surface | Separate approval and provisioning |
| Stop boundary | Stay inside owned or explicitly authorized systems | Stop before credential access, persistence, or production changes unless separately and explicitly authorized |
Read the Blue vs Red decision guide before selecting a model path.
Deployment Gates
1. Approval gate
Record the approval owner and the exact identity or service, OpenAI organization or project, offering, model, and surface. An application or identity check is not approval.
2. AWS gate
Confirm the selected model is available in the intended Region. Validate the AWS identity and Bedrock permissions using a non-production test. Model availability, quotas, billing, and service logs are AWS-side facts.
3. Runtime gate
For supported local ChatGPT Work and Codex surfaces, the client sends requests to Bedrock using AWS-native authentication. The OpenAI-hosted Responses API is not in that request path, and OPENAI_API_KEY is not the credential for this provider.
Use the Bedrock Mantle implementation guide for the provider, Region, and credential checks.
4. Authority gate
Write the permitted systems, task window, allowed actions, forbidden actions, and stop conditions. A model entitlement does not authorize access to a target or permission to modify production.
5. Operations gate
Start in an isolated environment with least privilege. Log requests and tool actions according to approved policy, require review for sensitive actions, and keep a rollback and incident path.
A Safe Pilot Sequence
- Choose a Blue workflow with a bounded repository or lab dataset.
- Confirm Daybreak and AWS access with the exact production identity, but test only in the isolated pilot environment.
- Begin with read-only analysis and a known-answer evaluation set.
- Measure evidence quality, false positives, latency, cost, and reviewer effort.
- Allow patch proposals before allowing patch application.
- Add a human confirmation gate for network, credential, deployment, or destructive actions.
- Re-review approval and scope before moving to Red or a wider target set.
Run the Daybreak AWS deployment readiness tool to turn these gates into a review record.
Frequently Asked Questions
Is OpenAI Daybreak publicly available in Amazon Bedrock?
No. OpenAI says Daybreak Blue and Red require enrollment in Daybreak Access. Approval and provisioning must cover the intended identity, organization or project, model, and surface.
Does Daybreak Blue approval include Daybreak Red?
No. OpenAI documents Daybreak Red as a separately approved and provisioned specialist offering. Blue approval or identity verification does not grant Red access.
Can I use an OpenAI API key with Bedrock Mantle?
Not for the Bedrock-backed local provider described by OpenAI. That path uses a Bedrock API key or the AWS SDK credential chain, and the OpenAI-hosted Responses API is not in the request path.
Does Bedrock model access authorize a penetration test?
No. AWS entitlement and model approval are separate from authorization to assess a system. Use Daybreak only on systems you own or are explicitly authorized to test, within the documented engagement scope.
Should most security teams start with Daybreak Red?
No. OpenAI recommends Daybreak Blue for most authorized defensive work. Red is a specialist option for separately approved advanced workflows.
Official Sources
- OpenAI: Daybreak models are now available on AWS
- OpenAI: Models and Trusted Access
- OpenAI: Use ChatGPT Work and Codex with Amazon Bedrock
- AWS: OpenAI models in Amazon Bedrock
Next Decision
- Compare Daybreak Blue and Red for the intended workflow.
- Verify Daybreak enrollment and AWS access.
- Plan security governance and operations before production.