A "Cyber-Permissive" Version of GPT-5.6 Sol
On Monday, August 11, 2026, OpenAI announced the release of GPT-5.6-Cyber, a new artificial intelligence model specialized in vulnerability research, penetration testing, and incident response. Built on the foundation of GPT-5.6 Sol, it is trained to improve its performance on specialized cybersecurity tasks — including the discovery of zero-day vulnerabilities and the development of exploit chains — while reducing the refusal rate in response to high-risk dual-use cyber solicitations.
“Built on GPT-5.6 Sol, it is trained to enhance its capabilities across several specialized cybersecurity tasks (e.g., finding zero-day vulnerabilities and developing exploit chains) and to reduce refusals on certain high-risk cyber tasks and dual-use,” states OpenAI.
GPT-5.6-Cyber succeeds GPT-5.5-Cyber, released in June 2026, and is part of a broader strategy called Daybreak, launched in May 2026. This initiative includes two distinct levels of access:
- Daybreak Blue: provides access to cutting-edge generalist models, such as GPT-5.6 Sol, with built-in safeguards suited for defensive security work. OpenAI specifies that this level “removes these safeguards to help defenders get more out of the model in real security tasks” such as incident detection, investigations, vulnerability management, and security assessments.
- Daybreak Red: it is through this new tier that GPT-5.6-Cyber is made available to companies for authorized vulnerability research, exploit validation, and security testing. This is the most permissive level.
A 95% Completion Rate on Advanced Cyber Tasks
To measure the reduction in refusals, OpenAI developed an internal assessment called Advanced Cybersecurity Completion Rate. This metric gauges how frequently the models respond to requests related to exploit chain development, bypassing authentication, privilege escalation, and other advanced cybersecurity scenarios.
The results are telling:
| Model / Access Level | Completion Rate |
|---|---|
| GPT-5.6-Cyber (Daybreak Red) | 95.0% |
| GPT-5.6 Sol (standard model) | 1.5% |
| GPT-5.6 Sol (Daybreak Blue) | 2.0% |
| GPT-5.5-Cyber (previous model) | 57.3% |
The leap is striking: GPT-5.6-Cyber completes 95% of high-risk cyber queries, compared to just 1.5% for GPT-5.6 Sol without specialized access, and 57.3% for its predecessor GPT-5.5-Cyber.
On the ExploitGym benchmark, GPT-5.6-Cyber outperforms both GPT-5.6 Sol and GPT-5.5-Cyber. The model also shows improvements in the discovery and calibration of the severity of novel zero-day vulnerabilities, thanks to specialized training.
Vulnerabilities Already Discovered, Including a Critical Flaw in V8
The concrete results of the model are already visible. OpenAI states that GPT-5.6-Cyber has identified several high-severity flaws, the most notable being CVE-2026-15903 (CVSS 3.1 score: 8.8 — High). This is a boundary read and write vulnerability in Google Chrome's V8 JavaScript engine, which could allow a remote attacker to execute arbitrary code inside a sandbox via a malicious HTML page. This flaw could be chained with another undocumented vulnerability, also discovered by the model, to escape the V8 heap sandbox.
Independent Confirmation
The details of this vulnerability are corroborated by several public sources. The National Vulnerability Database (NVD) from NIST confirms that CVE-2026-15903 corresponds to a boundary read and write in V8 of Google Chrome prior to version 150.0.7871.128, with a CVSS 3.1 score of 8.8 (High), vector AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H.
Google’s Chrome release blog provides a crucial detail: the flaw was reported by “OpenAI Codex Security (amyb)” on July 6, 2026. It was patched in the stable update on July 16, 2026 (Chrome 150.0.7871.128). This mention directly confirms that OpenAI's tools were behind the discovery of this vulnerability.
Beyond this flaw, OpenAI indicates that the model has also enabled the reporting of:
- At least 5 vulnerabilities in a popular mobile operating system, including a chain from an untrusted application to local privilege escalation
- 3 critical vulnerabilities in a popular database, including a remote access path leading to code execution
- Over 400 vulnerabilities that could lead to privilege escalation in the kernel of a popular operating system
Trusted Partners, but Safeguards Lifted
GPT-5.6-Cyber is already made available to a group of trusted partner clients, including: Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Palo Alto Networks, PwC, and Sophos. The stated goal is to help these companies identify and correct vulnerabilities before attackers can exploit them, bridging the “defensive gap.”
However, OpenAI explicitly acknowledges the inherent risks of such an approach:
“Models operating with reduced safeguards come with risks beyond standard model usage, whether misuse or misalignment. Despite these risks, we believe democratizing access to cutting-edge intelligence for defenders is crucial to accelerating and automating cybersecurity.”
The Context: AI as a Double-Edged Sword
This launch occurs in a landscape where AI is increasingly used by attackers. AI agents enable cybercriminals and state-sponsored hackers to outsource the tedious work necessary for planning and executing attacks, thus enhancing the efficiency and productivity of their operations. AI has also shortened the time between the disclosure of a vulnerability and its exploitation, with attackers relying on these tools to write what The Hacker News describes as “vibe exploits” — quickly generated exploits for newly disclosed flaws.
However, while AI systems have become highly proficient at finding and exploiting vulnerabilities, they remain noticeably less adept at fixing them. Research conducted by 1Password and cited by The Hacker News shows that reasoning models capable in the cyber domain, like ChatGPT 5.5 and Anthropic Claude Opus 4.8, struggle to fully patch a discovered vulnerability or avoid introducing new issues into their fixes:
- Success rate for generating a fix that entirely resolves the vulnerability (without altering application behavior): 26.0%
- Fixes resolving the vulnerability but altering the behavior of the application: 20.1%
- Fixes failing to resolve the vulnerability, introducing a new vulnerability, or both: 53.9%
These figures highlight that models excelling in discovering a broad range of vulnerabilities are only effective at patching a narrow subset. Worse, in many cases, fixes generated by LLMs introduce new bugs regardless of the resolution of the initial issue, thereby expanding the attack surface for malicious actors.
A Notable Limitation: Less Detailed Vulnerability Reports
GPT-5.6-Cyber is not superior in all respects. OpenAI acknowledges that the model performs worse than GPT-5.6 Sol on open tasks associated with discovering vulnerabilities in a code repository, developing a functional proof of concept, and submitting a quality vulnerability report. This regression is attributed to the model sometimes producing “shorter and less detailed vulnerability reports”.
Issues and Perspectives
The launch of GPT-5.6-Cyber marks a significant milestone in the race for AI applied to cybersecurity. By drastically reducing refusals on exploitation tasks, OpenAI is making a calculated bet: arming defenders with capabilities equal to or even surpassing those of attackers. The logic is that of an arms race where AI must bridge the gap between the time attackers take to exploit a vulnerability and the time defenders have to fix it.
But this bet raises fundamental questions. The democratization of access to models capable of developing exploit chains, even when limited to trusted partners, expands the risk surface in case of misuse or compromise. And as demonstrated by data from 1Password, the ability to find a flaw does not guarantee the ability to fix it correctly — a paradox that could ultimately exacerbate security technical debt rather than resolve it.
With AI already lowering the barrier to exploitation and accelerating vulnerability research, attackers are likely to cast a wider net on vulnerabilities disclosed in the future to find a foothold in enterprise networks. The question is no longer whether AI will transform cybersecurity, but who will make the best and fastest use of it.