OpenAI unveils GPT-6 Astra, first 'critical' risk classification for cybersecurity
The company describes the new model as the most capable ever built for software development, but acknowledges it is also the first system classified as critical risk for its autonomous computing capabilities.
OpenAI president Greg Brockman presented GPT-6 Astra as the result of years of research and investment, describing each of the company’s advances as built upon the previous one.
“It brings together years of research and sustained investment, every advance built on the one before” (transl. from Russian) — Greg Brockman, president of OpenAI
According to the company, the model stands out in particular in software development, with performance surpassing competitors in bug hunting and in executing terminal-based tasks. These are claims reported by OpenAI itself and not verified by third parties in the available excerpts: at present, no independent benchmarks are cited by the sources in support of these comparisons.
The most significant point, however, concerns safety: Astra is the first OpenAI model to receive a “critical” risk level in the cybersecurity domain, owing to its ability to autonomously find and exploit vulnerabilities in protected systems. It is the company itself that assigned this classification to the model, not an external certification body indicated in the available excerpts. This is a higher threshold than those assigned so far to previous OpenAI models, according to the company’s statement.
OpenAI has also acknowledged that the new system is harder to monitor externally compared with previous models, an admission that accompanies, in the company’s communication, the announcement of its most advanced capabilities. The available sources do not specify which monitoring tools were tested nor what the greater difficulty of external oversight actually consists of.
The picture that emerges is one of a company claiming a leap in capability while, at the same time, declaring an increase in the risk associated with that same capability. No independent assessments by third-party bodies regarding the risk classification assigned to the model appear in the available excerpts, nor are any timelines announced for a potential wider public release of Astra.
← Archive · Front page · Past editorials · Report an error · Original article (in Italian)