What is GPT-5.6-Cyber?

GPT-5.6-Cyber is OpenAI’s purpose-trained cybersecurity model available through Daybreak Red. Built on GPT-5.6 Sol, it is fine-tuned to improve performance on specialized security tasks like zero-day vulnerability discovery and exploit chain development while reducing refusal rates for authorized dual-use security work.

What is the difference between Daybreak Blue and Daybreak Red?

Daybreak Blue provides access to general-purpose frontier models like GPT-5.6 Sol with system-level guardrails removed to support defensive tasks like incident response, malware triage, and secure code review. Daybreak Red provides access to dedicated cybersecurity models like GPT-5.6-Cyber for authorized vulnerability research and exploit validation.

What is CVE-2026-15903 and how was it found?

CVE-2026-15903 is a high-severity vulnerability in Chrome’s V8 JavaScript engine discovered using GPT-5.6-Cyber. The JIT optimizing compiler incorrectly skipped safety checks when converting values to integers, allowing out-of-bounds memory reads/writes that could be chained with a JSPI stack escape to execute arbitrary code outside the V8 heap sandbox.

What are the access requirements for Daybreak models?

Daybreak access is restricted to approved individuals and organizations conducting authorized security work. Requirements include identity verification, account monitoring, legal attestations, sandboxed execution, and mandatory hardware security keys starting September 1, 2026.

Inside GPT-5.6-Cyber: OpenAI’s Dedicated Defensive | SPIDITS AI

Metric	Gemini 3.6 Flash	Gemini 3.5 Flash	Improvement	Evaluation Scope
Advanced Cybersecurity Completion Rate	95.0% (GPT-5.6-Cyber Red)	1.5% (GPT-5.6 Sol Default)	+93.5% completion rate	Measures responses to exploit-chain development, authentication bypass, and privilege escalation prompts.
Daybreak Blue Sol Completion Rate	2.0% (Daybreak Blue)	1.5% (Default)	+0.5% completion	Demonstrates that general-purpose Sol still refuses highly dual-use prompts even with reduced guardrails.
GPT-5.5-Cyber Baseline Comparison	95.0% (v5.6-Cyber)	57.3% (v5.5-Cyber)	+37.7% gain over predecessor	Addresses persistent refusal issues encountered by security researchers in earlier Cyber models.

Executive Summary & Strategic Rationale

On August 10, 2026, OpenAI officially published "Expanding Daybreak as the Cyber Defense Window Narrows," detailing a major evolution in its cybersecurity intelligence deployment strategy.

The threat landscape is undergoing a structural shift: malicious actors are increasingly utilizing artificial intelligence to conduct cyberattacks at unprecedented speed, scale, and autonomy. As these offensive AI capabilities proliferate, security defenders face a rapidly narrowing window of preparation.

To rebalance the defense-offense dynamic, OpenAI is expanding its Daybreak program by launching two distinct access tiers—Daybreak Blue and Daybreak Red—and introducing GPT-5.6-Cyber, its latest specialized model engineered specifically for advanced cybersecurity research and vulnerability discovery.

Program Architecture: Daybreak Blue vs. Daybreak Red

In production, general-purpose LLMs rely on automated system-level guardrails to block potential cyber misuse. However, these guardrails frequently block legitimate defensive tasks such as incident response, malware analysis, and security assessments. Daybreak solves this through a dual-tiered architecture:

Access Tier	Underlying Base Model	Key Functional Scope	Refusal Profile & Safeguards
Daybreak Blue	General-purpose GPT-5.6 Sol	Vulnerability discovery, secure code review, malware analysis, incident response, patch validation	System-level guardrails removed for defensive workflows; refuses highly dual-use prompts
Daybreak Red	Purpose-trained GPT-5.6-Cyber	Authorized vulnerability research, exploit validation, deep red teaming, zero-day discovery	Purpose-trained to reduce refusals on high-risk, dual-use security prompts

Deep-Dive: GPT-5.6-Cyber & Refusal Rate Calibration

Even when general-purpose models like GPT-5.6 Sol run under Daybreak Blue with system-level guardrails removed, they continue to refuse highly dual-use prompts (e.g., pen-testing production environments or bypassing authentication logic).

To address this friction, GPT-5.6-Cyber was trained on top of GPT-5.6 Sol specifically to reduce unnecessary refusals for vetted security professionals while improving technical reasoning over complex exploit constraints.

Advanced Cybersecurity Completion Rate Benchmark

To evaluate model compliance on dual-use security tasks, OpenAI established an internal evaluation measuring response rates to prompts involving exploit-chain development, authentication bypass, privilege escalation, and macOS Keychain/Chrome cookie decryption:

Model Variant	Access Tier	Completion Rate (%)	Refusal Behavior
GPT-5.6 Sol	Safeguards Enabled	1.5%	Refuses 98.5% of dual-use security prompts
GPT-5.6 Sol	Daybreak Blue	2.0%	Refuses 98.0% of dual-use security prompts
GPT-5.5-Cyber	Daybreak Red	57.3%	Moderate refusal reduction over earlier releases
GPT-5.6-Cyber	Daybreak Red	95.0%	Permissive (95.0% completion rate for authorized work)

While standard GPT-5.6 Sol refuses 98.5% of dual-use prompts even under Daybreak Blue, GPT-5.6-Cyber completes 95.0% of authorized security requests, eliminating persistent refusals that previously stalled legitimate research.

Evaluation Metrics: ExploitGym, ExploitBench & Zero-Day Evals

OpenAI evaluated GPT-5.6-Cyber across several rigorous cybersecurity benchmark suites:

ExploitGym: Evaluates whether autonomous agents can turn known software vulnerabilities into working exploits achieving arbitrary code execution in controlled, isolated environments. GPT-5.6-Cyber significantly outperformed both GPT-5.6 Sol and GPT-5.5 Cyber.

Zero-Day Discovery Eval: Measures a model's ability to analyze open-source code repositories, identify novel zero-day flaws, generate proof-of-concept (PoC) exploits with maximum impact, and author calibrated technical reports. GPT-5.6-Cyber (Daybreak Red) achieved higher average discovery quality than GPT-5.6 Sol (Daybreak Blue).

Vulnerability Discovery & Report Writing: Evaluates open-ended repo auditing. On this benchmark, GPT-5.6 Sol scored slightly higher than GPT-5.6-Cyber because GPT-5.6-Cyber occasionally generated shorter, less verbose technical write-ups.

ExploitBench: Tests developing complex V8 JavaScript engine vulnerabilities into full sandbox-escaping exploits under strict turn limits. In standard 300-turn evaluations (with V8 sandbox enabled), GPT-5.6 Sol (Daybreak Blue) exhibited superior token efficiency. When expanded to 600 turns, the performance gap between Sol and Cyber narrowed.

Real-World Findings: CVE-2026-15903 & Kernel Vulnerabilities

Beyond lab benchmarks, OpenAI researchers utilized GPT-5.6-Cyber to conduct real-world security audits across major software codebases, resulting in critical vulnerability disclosures:

1. Chrome V8 Sandbox Escape (CVE-2026-15903)

GPT-5.6-Cyber uncovered two previously unknown vulnerabilities in Chrome's V8 JavaScript engine that could be chained together to corrupt memory and escape the V8 heap sandbox:

The Mechanism: In V8's optimizing compiler (JIT), an incorrect type conversion logic skipped a safety check when casting values to integers, causing undefined values to evaluate as unexpectedly large numbers.

Exploitation Path: When used as an array index, the compiler omitted bounds checking. An attacker could read or overwrite out-of-bounds memory, achieving arbitrary read/write inside Chrome's heap sandbox.

Sandbox Escape: Combined with a second vulnerability discovered by GPT-5.6-Cyber—a JSPI stack escape—the chain enabled native code execution outside the V8 sandbox.

Remediation: OpenAI reported the findings via coordinated disclosure, and Google patched the flaw under CVE-2026-15903.

2. Mobile, Database & OS Kernel Audits

Mobile OS Kernel: Discovered at least 5 vulnerabilities in a popular mobile operating system, including an untrusted application privilege escalation chain.

Database Infrastructure: Identified 3 critical vulnerabilities in a widely used database system, providing remote code execution (RCE) paths.

Operating System Kernel: Uncovered over 400 vulnerabilities capable of local privilege escalation across a major OS kernel.

Partner Validation & Enterprise Deployment

Early access to GPT-5.6-Cyber was provisioned to selected cybersecurity partners—including SpecterOps, SentinelOne, and Palo Alto Networks—to evaluate model performance across active defensive research workflows.

Security teams at SpecterOps reported that the cyber-permissive model significantly accelerated vulnerability research and complex state tracking. By reducing false-positive refusals during authorized security assessments, researchers completed complex exploit validations within hours that previously required weeks of manual auditing.

Preparedness Framework & Hardware Security Controls

Under OpenAI’s internal Preparedness Framework, GPT-5.6-Cyber was assessed as reaching the High capability threshold for cybersecurity, remaining safely below the Critical threshold (defined as unassisted zero-day discovery and autonomous infrastructure exploitation without human intervention).

(Note: OpenAI explicitly confirmed that GPT-5.6-Cyber was not involved in the previously reported Hugging Face security incident).

Mandatory Operational Safeguards:

Hardware Security Keys: All individual accounts in Daybreak are required to adopt physical hardware security keys (YubiKey / WebAuthn) starting September 1, 2026.

Codex Auto-Review Mode: UI defaults strongly encourage switching from full-access execution to auto-review mode, which evaluates actions requiring elevated permissions prior to execution.

Air-Gapped Sandboxing: High-risk workflows must run in isolated environments without direct access to production systems or the open internet.

Scoped Permission Profiles: Organizations must define granular permission boundaries for agent tool calls.

Key Strategic Takeaways for Defense Teams

Adopt Permissive Defensive AI: Relying on default LLMs with strict safety guardrails leaves security researchers at a disadvantage against unrestricted threat actors.

Accelerate Remediation Pipelines: With models finding over 400 kernel vulnerabilities and 95.0% completion rates on exploit chains, defensive teams must automate patch verification.

Mandate Hardware-Bound Authentication: Deploying frontier cyber models requires strict identity verification, YubiKey hardware tokens, and sandboxed execution loops.

Inside GPT-5.6-Cyber: OpenAI’s Dedicated Defensive Frontier Model and the Daybreak Expansion