An unannounced name in an AI coding tool's list
The latest version of the desktop app for "Devin," the AI coding tool from U.S.-based Cognition, contains an unannounced model name, "astra-cyber-preview," @aitrackerbot, an X account that tracks the appearance of AI models, reported on Oct. 18 (Oct. 19 Japan time). Its settings allowed reasoning depth (how much the AI thinks before answering) to be set up to the top "ultra" level. Judging from the name, it appears to be a preview of a version of OpenAI's top model, "GPT-6 Astra," tailored for cybersecurity. OpenAI has not announced it, and Cognition has not commented.
Devin is an AI coding service in which the AI plans its own steps and writes programs. It has been quick to adopt new OpenAI models, making GPT-6 Astra available in its desktop and command-line versions on Sept. 3.
Models with strong cyber capabilities offered by approval only
OpenAI provides models with high cyber capabilities that could also be used for attacks only to approved experts and organizations. In June it launched "Daybreak," an initiative to support defenders, and in August it split it into two tiers. "Daybreak Blue" provides general-purpose models with safeguards for defensive work, while "Daybreak Red" provides models specially trained for investigating vulnerabilities (software flaws) and verifying attack procedures. Access requires identity verification, monitoring and agreement to usage restrictions.
OpenAI said "GPT-5.6-Cyber," released in August, complied with 95.0% of requests from defensive experts. The previous "GPT-5.5-Cyber" complied with 57.3%. The general-purpose "GPT-5.6 Sol," in its standard offering, complies with only about 1.5% of similar requests. These are the company's own figures, without external verification.
"GPT-6 Cyber" reported in September
U.S.-based Fortune reported in September that OpenAI was offering "GPT-6 Cyber" on a trial basis to some customers through Daybreak Red. It also reported that the model might be shown at the "DevDay" developer conference on Sept. 29, but Reuters said it could not independently confirm this. Whether "astra-cyber-preview" is the same model as the one in that report is unknown.
Companies are cautiously expanding the availability of cyber-focused models. Google announced on Oct. 29 that it would provide its top model, "Gemini 4 Carbon," first to cyber defenders. xAI is also believed to be preparing a category in its API that allows offensive and defensive content only for approved keys.
