OpenAI has introduced GPT -5.6-Cyber , a specialized model for cybersecurity tasks, aimed at vulnerability research, exploit validation, and security testing. Its availability through Daybreak Red comes with fewer denials on certain high-risk requests.
According to OpenAI's official announcement, the new model is based on GPT-5.6 Sol and has been trained to detect unknown vulnerabilities, develop exploit chains, and perform advanced security research. The company emphasizes that access is reserved for authorized defense researchers.
See also: OpenAI Astra: Pause on most powerful AI model due to critical cybersecurity capabilities
GPT-5.6-Cyber and the possibilities in research
Daybreak is divided into two tiers. Blue offers general-purpose models, such as GPT-5.6 Sol, with configurations for authorized defense, secure code review, malware analysis, and incident response. Red is aimed at teams looking for vulnerability research, exploit validation, and penetration testing.
OpenAI created an internal assessment called Advanced Cybersecurity Completion Rate. On requests involving exploit chain development, authentication bypass, and privilege escalation, GPT-5.6-Cyber running Daybreak Red achieved a 95% completion rate. GPT-5.5-Cyber achieved 57.3%, while GPT-5.6 Sol with standard protections achieved just 1.5%.
The company also claims that the model identified two previously unknown issues in Chrome's V8 JavaScript engine. The findings could be combined to cause memory corruption and a shell script exit, and Google has patched them and assigned the related chain CVE-2026-15903.
In the same announcement, OpenAI lists at least five vulnerabilities in a popular mobile operating system, three critical vulnerabilities in a popular database, and more than 400 issues that could lead to privilege escalation in the operating kernel. It does not name the specific products, as coordinated disclosure processes are ongoing.
See also: AI sandbox: Testing reaches real systems
The limits and risks of GPT-5.6-Cyber
The reduction of protective barriers is the most sensitive point of the announcement. OpenAI recognizes that a model with fewer restrictions can be misused or have alignment problems. That's why it recommends isolated environments, without access to production systems or the open internet.
The available data is not all successes. In an internal test for finding vulnerabilities and writing a technical report, GPT-5.6-Cyber performed worse than GPT-5.6 Sol, because it produced shorter and less detailed reports. In another test, GPT-5.6 Sol was more efficient when the task was limited to 300 steps.
Access to both levels of Daybreak is limited to approved individuals and organizations. The process includes identity verification, account review, monitoring, permitted use statements, and legal attestations. Starting September 1, 2026, individual accounts will be required to use hardware security keys.
OpenAI also suggests enabling an automated review feature in Codex so that actions that require elevated privileges are reviewed before they are executed. Organizations can set their own policies, but the company notes that the technology is not a substitute for expert review, especially when a test could impact data or critical services.
The announcement presents the results as internal measurements and does not provide independent verification of the scores or full details of the testing methodology. OpenAI says it will release a system card with more evaluations later. Until then, the performance should be treated as manufacturer claims and not as universal proof of security.
The choice between the two access levels depends on the nature of the work. For code review, incident analysis, and fix validation, Blue is recommended as a more limited starting point. Red is for research teams that have documented authorization and can demonstrate that their testing remains within predefined boundaries.
🔒 Protect your privacy with Proton VPN
Swiss VPN from the creators of Proton Mail — strict no-logs policy, strong encryption, and built-in NetShield that blocks ads, trackers, & malware.
- ✔ No-logs, based in Switzerland (except 14-Eyes)
- ✔ NetShield: blocks ads, trackers & malicious domains
- ✔ Covers all devices — free version available
The link is an affiliate link — SecNews may receive a commission at no additional cost to you. It does not affect the independence of our article writing.
GPT -5.6-Cyber should not be treated as a standalone security auditor. Its value lies in accelerating iterative checks and formulating hypotheses, while the final decision on whether a finding is real, serious, and fixable remains the responsibility of the human team.
See also: Cloudflare Kitesurf: Browser for AI agents
The position of the SecNews technical team
GPT -5.6-Cyber shows that models can speed up vulnerability discovery, but high performance in finding a problem does not equal secure patching. Every finding requires human confirmation, coordinated disclosure, and verification that the patch does not create a new risk.
The SecNews technical team considers it critical to use segmented labs, limited permissions, and full logging of actions before any testing on real infrastructure. Access to stronger cybersecurity tools can only strengthen defenses when accompanied by clear authorization and constant human oversight.
