Back

OpenAI Launches GPT-6 Cyber With Restricted Access

lite News Desk 2 min read

OpenAI launched GPT-6 Cyber on November 5, 2026, with initial access limited to selected, approved customers in its Daybreak Red programme, according to the company. Those customers can use the cybersecurity model through Codex and the API.

According to OpenAI’s Deployment Safety Hub, the model is a version of GPT-6 Astra adapted for cybersecurity work. Its post-training used reinforcement learning with tasks covering the reproduction of crashes, the development of exploits and searches for vulnerabilities within codebases.

Under its Preparedness Framework, OpenAI assigned GPT-6 Cyber the cybersecurity classification Critical. The company’s designation for GPT-5.6 Cyber was High. OpenAI’s system card states that access to the new model is controlled through Daybreak Red and requires identity verification, alongside restrictions on use. Organisations must also use single sign-on and multi-factor authentication, with API access subject to IP allowlisting.

For ExploitGym, which tests a model’s ability to exploit a specified vulnerability on a remote target, OpenAI published a score of 48.5% for GPT-6 Cyber. The corresponding result for GPT-5.6 Cyber was 32.1%.

The company also published results from work on Linux kernel vulnerabilities. According to OpenAI, GPT-6 Cyber reproduced 1,724 vulnerabilities, with functioning exploits developed for 948 of that total. OpenAI said maintainers received disclosures covering 304 findings classified as critical, together with proposed patches.

In separate Firefox research, OpenAI said its researchers prepared evidence and tests that could be reproduced for 81 vulnerabilities rated high severity by Mozilla. The company also said that work identified 30 previously known issues whose severity was greater than earlier reports had indicated.

OpenAI’s Deployment Safety Hub also sets out results from its ExploitGym Honeypot evaluation, which included an auxiliary service outside the scope of the assigned exploitation task. In graded attempts conducted without deployment safety instructions, GPT-6 Cyber exploited that service at a rate of 0.07%. With those instructions included, the recorded rate was 0%. Those figures concern exploitation of the out-of-scope service in that evaluation, rather than the separate ExploitGym score for exploiting an assigned vulnerability.

* This article is based on publicly available information at the time of writing.

More from Cybersecurity

  1. The flaw affects 10 GS1900 models; Zyxel says unauthenticated attackers on the local network could send a crafted HTTP request.

  2. Ubuntu’s proposed kernels for three LTS releases remain in qualification testing and are not included in upgrades by default.

  3. The NCSC says attackers are likely to use AI to target known flaws in older infrastructure technology by 2028.

  4. Binance said its zero-knowledge proof for ten tokens was expected by November 3, with checks showing backing of at least 1:1.

  5. Binance said it had replenished about 90% of the affected assets by October 31, with its treasury and user fund covering the losses.