OpenAI unveils Astra model with expanded cyber capabilities
On Thursday, OpenAI unveiled a new artificial intelligence model, Astra, which the company calls its most powerful and capable development to date. According to OpenAI, the model opens a “new frontier” in working with computers and browsers and performs tasks with unprecedented speed, accuracy, and security.
Astra initially became available to clients of OpenAI’s Daybreak cybersecurity program. Over the next week, the company also plans to provide access to users of the paid Pro, Plus, Enterprise, and Business plans, as well as through the API. TechCrunch reports on the launch and the model’s stated capabilities.
Cybersecurity and programming
OpenAI President Greg Brockman called Astra the company’s most intelligent model and the one most aligned with users’ intentions. Such alignment refers to the model’s tendency to act in accordance with a user’s requests or in their interests. According to Brockman, Astra combines the results of years of research and could change the nature of work that people delegate to artificial intelligence.
OpenAI said it tested Astra against a number of security benchmarks. The company claims that the model can detect and develop zero-day exploits, which could help security specialists identify and eliminate vulnerabilities. OpenAI also calls Astra its best model for software engineering and cites test results in which, according to the company, it outperforms other models, including OpenAI Sol and Anthropic Fable, in finding bugs, completing terminal tasks, and answering queries about codebases.
More current news is available on the UA.News Telegram channel Telegram.
Reasoning transparency and AGI
At the same time, Astra has drawn attention because of its use of the opaque recurrence technique. It may make it more difficult to monitor the model’s chain of reasoning — a process that allows researchers to examine how and why artificial intelligence made a particular decision.
OpenAI Chief Scientist Jakub Pachocki said that control over the reasoning process is an important form of oversight, but as models’ capabilities grow, their observability becomes more difficult. He suggested that more powerful models may perform more complex tasks with fewer language tokens or without them, making such actions harder to track.
Answering a question about artificial general intelligence, Brockman said that there is no longer a contractual criterion for AGI. He was referring to a previous condition of OpenAI’s agreement with Microsoft under which the partnership was to end after AGI was achieved. According to Brockman, the concept of AGI has become more of a “mission or spiritual concept,” and readers should decide for themselves whether Astra meets this definition. At the same time, he added that he personally believes this level has already been achieved.
