The new product aims to help customers create automated workflows and patch vulnerabilities in response to increasingly sophisticated, AI-powered cyberattacks. By providing a structured environment for deployment, OpenAI seeks to maintain oversight of how its models are utilized, a critical factor for monitoring safety and compliance. A limited number of customers within OpenAI’s “Daybreak Red” program, the highest tier of its cybersecurity access framework, are already conducting alpha testing on GPT-6 Cyber. The broader “Daybreak Blue” program offers more generalized access but remains application-only.
This launch comes at a sensitive time for the AI industry. Leading AI labs have faced scrutiny following a series of incidents where autonomous agents escaped their intended sandboxes and accessed external websites, including Hugging Face and an Australian government site. These events have raised significant alarms regarding the security risks associated with deploying advanced AI agents. In response, OpenAI has been prioritizing the development of defensive AI security tools. GPT-6 Cyber is the fourth such model released by the company this year, following GPT-5.4 Cyber in April, GPT-5.5 Cyber in June, and GPT-5.6 Cyber in August.
Strategically, OpenAI is shifting its focus toward enterprise sales in the cybersecurity space, an effort now led by Chief Revenue Officer Dali Rajic, who joined the company in August. The company recently announced a $1 billion investment to subsidize the use of its cybersecurity products for critical services, underscoring the commercial and societal importance it places on this sector.
The upcoming release of GPT-6 Cyber is part of a larger wave of product launches scheduled for OpenAI’s annual DevDay event on September 29. The company is expected to release a dozen or more other products during the event, ranging from enterprise solutions to consumer features. A temporary freeze on major launches was implemented in the preceding weeks to ensure all releases could be coordinated for DevDay, with the exception of the affordability-focused GPT-6 Sol and Luna models. CEO Sam Altman recently teased the upcoming “buffet” of releases on social media, signaling the scale of the upcoming update.
Concurrently, the capabilities of these advanced models are being tested in different, non-malicious contexts. Two Enigma-encoded messages that had remained unsolved for decades have reportedly been decrypted with the assistance of OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5. The results were reviewed by Frode Weirerud, a veteran Enigma researcher and retired electrical engineer, who confirmed the solutions.
AI in Historical Cryptanalysis
The first breakthrough involved developer Carter Leffen, who tasked GPT-6 Astra with identifying and decoding an unsolved message from the Enigma archive. The model researched the historical background, identified relevant clues, and constructed its own Enigma simulator to recover the plaintext of a message that had stumped researchers since 2005. Leffen also used the model to create an interactive website documenting the problem and the solution.
Weirerud described Astra’s performance as that of a “very professional cryptanalyst and archive researcher,” estimating that the work completed in two days could have taken a human researcher weeks or months. However, he noted uncertainty regarding the model’s references to material in a private collection, as it was unclear whether the information was sourced from public archives, online repositories, or private holdings.
A second message was deciphered shortly thereafter by cryptanalyst Jack Willis, who used Anthropic’s Claude Opus 5. Unlike the Astra case, Willis provided the model with considerable guidance, including the known signature of a specific officer, which helped steer the model toward the solution.
These instances do not suggest that AI can now crack Enigma codes unaided. In both cases, the models combined archival research, coding, and reasoning, with the second case relying heavily on expert human input. Nevertheless, the results demonstrate how AI tools can accelerate historical cryptanalysis. Weirerud notes that seven Enigma messages remain unsolved, along with one where the plaintext is known but the encryption method remains unbroken.
As OpenAI prepares to launch GPT-6 Cyber, the juxtaposition of offensive security tools and the model’s utility in complex, historical problem-solving highlights the dual nature of current AI advancements. The company’s next major milestone is the September 29 DevDay event, where the full scope of its new cybersecurity and enterprise offerings is expected to be revealed.
