OpenAI has expanded its Daybreak program with two access tiers and introduced GPT‑5.6‑Cyber, a model specifically trained for cybersecurity tasks. The company frames the move as a response to a rapidly changing threat landscape in which attackers increasingly use AI to mount faster, larger-scale, and potentially autonomous cyberattacks; defenders therefore need access to frontier intelligence under controlled conditions.
Daybreak access tiers
OpenAI offers two Daybreak levels for approved defenders:
- Daybreak Blue: removes some system‑level guardrails so defenders can use the model more effectively for real‑world security tasks such as incident detection and response, investigations, vulnerability management, and security assessments.
- Daybreak Red: provides access for more advanced, higher‑risk work; this tier is how organizations can access GPT‑5.6‑Cyber.
Access is restricted to approved individuals and organizations. OpenAI controls access through identity verification, account security, monitoring, approved‑use restrictions, and legal attestations.
What is GPT‑5.6‑Cyber?
GPT‑5.6‑Cyber is built on GPT‑5.6 Sol but is additionally trained to improve performance on specialized cybersecurity tasks—examples include discovering zero‑day vulnerabilities, developing exploit chains, and other advanced, dual‑use activities. The model is intended to reduce refusal rates on certain high‑risk prompts so that authorized defenders can carry out legitimate security research and defensive operations more effectively.
Performance and internal evaluations
OpenAI ran multiple internal evaluations to measure the model’s capabilities:
-
Advanced Cybersecurity Completion Rate: an internal metric that measures how often models respond to requests involving exploit‑chain development, authentication bypass, privilege escalation, and similar advanced scenarios. On this metric GPT‑5.6‑Cyber completes 95.0% of such requests, compared with 1.5% for GPT‑5.6 Sol and 2.0% for GPT‑5.6 Sol used with Daybreak Blue. The prior GPT‑5.5‑Cyber completed 57.3%.
-
ExploitGym: evaluates whether agents can turn known vulnerabilities into working exploits that achieve arbitrary code execution in controlled environments. GPT‑5.6‑Cyber outperformed GPT‑5.6 Sol and GPT‑5.5‑Cyber on this benchmark.
-
Zero‑day discovery benchmark: on a dataset where models received the current release of an open‑source repository and were asked to generate maximum‑impact proof‑of‑concept exploits plus technical writeups, GPT‑5.6‑Cyber (Daybreak Red) outperformed GPT‑5.6 Sol (Daybreak Blue) in terms of severity and impact due to its specialized training.
-
Vulnerability Discovery and Report Writing: models were tasked with finding vulnerabilities in repos with a known flaw and producing proof‑of‑concepts and reports. Both GPT‑5.6 Sol and GPT‑5.6‑Cyber improved over GPT‑5.5‑Cyber, but GPT‑5.6‑Cyber produced shorter, sometimes less detailed reports and therefore performed worse than GPT‑5.6 Sol on this evaluation.
-
ExploitBench (V8 exploitation): this harder evaluation tests turning a V8 vulnerability into a full exploit under stronger defenses and with less information. In the standard 300‑turn setting GPT‑5.6 Sol (Daybreak Blue) solved tasks more token‑efficiently and performed best; extending the setting to 600 turns narrowed the performance gap.
OpenAI notes that GPT‑5.6‑Cyber generally uses a larger reasoning budget and higher token usage than GPT‑5.6 Sol.
Real‑world research and partner feedback
OpenAI provided early access to GPT‑5.6‑Cyber to a set of trusted customer partners. According to feedback, the model materially improved specialist vulnerability‑research workflows by reasoning more accurately about exploit constraints, tracking complex state better, and completing work in hours or days that earlier models had not resolved after weeks.
The OpenAI team also used GPT‑5.6‑Cyber to study selected software projects. Using the model they investigated V8, the JavaScript engine in Chrome, and discovered two previously unknown vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox. The findings were validated by researchers and reported to Google through coordinated disclosure; Google fixed the issue under CVE‑2026‑15903.
CVE‑2026‑15903 is classified as a high‑severity vulnerability in V8’s optimizing compiler. The compiler skipped a safety check when converting values to integers, allowing undefined values to become unexpectedly large numbers. If such a number is used as an array index, the compiler may omit the usual bounds check, enabling an attacker to read or overwrite memory belonging to other objects and potentially achieve arbitrary code execution inside Chrome’s sandbox. Escaping the heap sandbox typically requires chaining with another vulnerability, which GPT‑5.6‑Cyber also identified.
OpenAI also reports finding high‑severity issues across other software projects, from popular databases to mobile operating systems, and says it is working with Daybreak partners and open‑source communities to disclose and remediate those flaws.
Safety assessment and deployment controls
Under OpenAI’s Preparedness Framework, GPT‑5.6 Sol was assessed as reaching the High level for cybersecurity capability but below the Critical threshold. GPT‑5.6‑Cyber was also evaluated and determined to reach High but not Critical. The company states that the model improved on some specialized cyber tasks but not enough to cross the Critical threshold.
OpenAI warns that models run with reduced safeguards carry additional risks from misuse or misalignment. Nevertheless, it argues that providing defenders with advanced AI tools—under gated, monitored conditions—is important to accelerate defensive capabilities.
Recommended practices for Daybreak include customizing organizational review policies, restricting use to approved workflows, and monitoring. OpenAI recommends that most defender teams start with Daybreak Blue; teams performing advanced vulnerability research, exploit development, or red‑teaming can apply for Daybreak Red.
OpenAI says it will publish a detailed system card and further evaluations for GPT‑5.6‑Cyber at a later date.
Availability
Daybreak Blue and Daybreak Red are available to approved individuals and organizations conducting authorized work; organizations can apply at openai.com/daybreak/partners.
