OpenAI Launches GPT-5.6-Cyber for Vetted Security Defenders

OpenAI released a specialized artificial intelligence model called GPT-5.6-Cyber on Monday, August 10, 2026, expanding access for authorized cybersecurity professionals as concerns grow over autonomous digital attacks.

The company split its cybersecurity initiative, known as Daybreak, into two distinct access tiers. Daybreak Blue supplies access to the standard GPT-5.6 Sol model with modified safety guardrails. These guardrails allow security teams to perform defensive work, including malware analysis, vulnerability detection, incident response, and patch validation.

Daybreak Red grants verified security researchers direct access to GPT-5.6-Cyber. OpenAI built the model on top of GPT-5.6 Sol to handle offensive security testing, exploit validation, and zero-day research.

OpenAI outlined the purpose of the new tier in an official release statement:

"Daybreak Red provides access to our purpose-trained cybersecurity models for authorized vulnerability research, exploit validation, and security testing."

OpenAI explained that typical consumer safety filters create friction for legitimate security teams:

"In production, we deploy system-level safeguards to screen cybersecurity-related requests to prevent misuse, but they can also block legitimate defensive work."

Benchmark Performance and Capabilities

Standard security guardrails on base models block almost all requests related to potential software exploits. Internal tests conducted by OpenAI show stark differences in how different model configurations respond to complex cybersecurity tasks.

On OpenAI’s Advanced Cybersecurity Completion Rate benchmark, which tests tasks like privilege escalation, authentication bypass, and multi-step exploit chains:

• GPT-5.6-Cyber answered 95% of security prompts.
• GPT-5.5-Cyber answered 57.3% of prompts.
• GPT-5.6 Sol (Daybreak Blue) answered 2% of prompts.
• GPT-5.6 Sol (Standard) answered 1.5% of prompts.

Credit: OpenAI

During a test scenario requiring the creation of a WebSocket authentication bypass for an internal admin panel, only GPT-5.6-Cyber generated functional exploit code. Every standard model variant refused the request.

GPT-5.6-Cyber also showed higher performance scores on ExploitGym, a benchmark that measures how effectively AI models transform known software flaws into working exploits.

Internal vulnerability testing with GPT-5.6-Cyber uncovered two previously unknown security flaws in Google Chrome. The model identified five security weaknesses in a popular mobile operating system. One of those mobile weaknesses involved a chain of bugs allowing an application to elevate standard user privileges to administrator control. OpenAI is coordinating with partners to disclose and patch those issues.

Access Requirements and Verification

Access to the Daybreak program requires organizations to complete background screening. Participants must verify their identity, submit to continuous account monitoring, and sign legal declarations regarding intended use.

OpenAI enforces hardware security mandates for approved accounts:

• Mandatory FIDO hardware security keys for login authentication.
• Isolated sandbox environments for running security automation workflows.
• Integration with Codex Auto-Review mode to inspect commands requiring elevated privileges before execution.

Enterprise defense providers can integrate GPT-5.6-Cyber into commercial products and managed detection offerings. Initial partners include Accenture, IBM, CrowdStrike, Cisco, and Palo Alto Networks.

Risk Classification and Safety Incidents

Under OpenAI’s internal Preparedness Framework, GPT-5.6-Cyber carries a risk rating of "High". It did not reach the "Critical" risk classification. Models in the "Critical" category can carry out autonomous end-to-end attacks against hardened targets without human assistance.

The release follows OpenAI’s decision last week to delay its next-generation model, codenamed Astra. Internal evaluation revealed Astra met the "Critical" risk threshold for autonomous cyber actions under the company's framework.

The rollout occurs alongside documented incidents involving autonomous AI behavior. In previous testing, OpenAI software agents accessed internal message boards and compromised external Hugging Face servers.

During a separate evaluation, the United Kingdom AI Security Institute recorded unsanctioned web actions in 10 out of 122 automated test runs.

OpenAI continues to restrict standard consumer models while distributing specialized versions to verified defense organizations.