OpenAI
OpenAI brings Daybreak Red cyber model to AWS GovCloud
Daybreak Red, OpenAI's GPT-5.6 Cyber model for authorized security work, is now on Amazon Bedrock in AWS GovCloud (US-East) for approved customers.
OpenAI’s Daybreak Red, a cybersecurity model built for authorized vulnerability and exploit research, is now available in a second AWS Region, AWS GovCloud (US-East). Approved customers can reach it through Amazon Bedrock.
AWS announced the Region in a What’s New post dated Oct. 30, 2026. OpenAI listed the same release in its news feed Nov. 4.
Daybreak Red Is GPT-5.6 Cyber for Authorized Security Work
Daybreak Red is GPT-5.6 Cyber, and AWS said it is built for authorized vulnerability research, exploit reproduction and mitigation development. It gives verified defenders more room with fewer refusals for this work, along with stronger identity verification, monitoring and access controls.
It is one half of Daybreak, OpenAI’s cyber defense initiative. The other half, Daybreak Blue, runs GPT-5.6 Sol with safeguards calibrated for defensive work. AWS said in an Aug. 11 blog post that Blue is the right starting point for most security teams and Red is for advanced tasks. That post introduced both models on Bedrock in US East (Ohio).
Access Is Limited to Approved Customers
Customers must enroll in Daybreak access from OpenAI. They contact OpenAI or their AWS account team about eligibility and, once approved, work with the account team to enable the model. OpenAI’s feed describes it as available to “approved customers in AWS GovCloud (US-East) through Amazon Bedrock for authorized cybersecurity work.”
Inference Data Stays Out of Training
Inference data is not used to train models, and customers do not need to share Daybreak Red data with OpenAI, AWS said. The model runs on Bedrock’s next-generation inference engine with “zero-operator access enforced at the chip.”
In the Aug. 11 post, AWS said zero-operator access means even AWS operators cannot see prompts and completions during inference. Traffic that a classifier flags is retained by AWS for up to 30 days for automated abuse detection. Customers can request zero data retention, which means AWS keeps no copy of that traffic, through their AWS account team.
The Model Card Sets the Limits
AWS’s Bedrock model card lists a launch date of Aug. 12, 2026 and a 1,050,000-token context window. The model takes image and text input and returns text. It does not support fine-tuning and works only on the bedrock-mantle endpoint, not bedrock-runtime.
OpenAI Says Researchers Found Two Chrome Flaws With It
According to OpenAI, security researchers used GPT-5.6 Cyber through Daybreak Red to find two previously unknown vulnerabilities in V8, the JavaScript engine in Chrome. AWS relayed the account in its Aug. 11 post. Chained together, the flaws could enable memory corruption and a heap sandbox escape. The first was fixed as CVE-2026-15903.
Our View
AWS’s own description of Daybreak Red holds two statements side by side. No AWS operator can read prompts and completions while the model runs, and AWS keeps traffic its classifier flags for up to 30 days. The classifier is the automated filter AWS uses to detect abuse.
The Aug. 11 post lists proprietary source code, unpatched vulnerability details and live telemetry among the inputs these workloads send the model. For security teams, and for agencies and contractors working in GovCloud, that material falls under the first statement while the model runs and under the second whenever the classifier flags the traffic.
The remedy is a zero data retention request to the same AWS account team that enables access. The protection holds only if AWS grants it.