ISLAMABAD – The company says Astra can identify and develop zero-day exploits to help security teams find vulnerabilities. Astra has also been presented as OpenAI’s strongest model for software engineering.
Its use of opaque recurrence has raised concerns about how researchers can monitor its reasoning. OpenAI President Greg Brockman said users could decide for themselves whether Astra meets their definition of artificial general intelligence (AGI).
Read more: Here are smartphones arriving in September: Details
OpenAI unveils Astra
OpenAI has launched Astra, describing it as its most powerful and capable artificial intelligence model to date.
The company said Astra sets a new standard for computer and browser use. It also highlighted improvements in speed, accuracy and safety.
OpenAI President Greg Brockman described Astra as the company’s most intelligent and aligned model so far.
According to Brockman, Astra brings together years of research and represents a shift in the kind of work that users can delegate to AI systems.
Astra availability
Astra is initially available to OpenAI customers using Daybreak, the company’s cybersecurity programme.
OpenAI plans to expand access over the following week to customers on its paid Pro, Plus, Enterprise and Business plans.
The company also plans to make Astra available through the OpenAI API, allowing developers to integrate the model into their own applications and services.
OpenAI highlights cybersecurity capabilities
Cybersecurity is a major focus of Astra. OpenAI said it tested the model against several security benchmarks and introduced additional safeguards to reduce potential risks.
The company said Astra can identify and develop zero-day exploits. Such capabilities could help security teams discover vulnerabilities and fix them before attackers exploit them.
The claims come as the AI industry faces growing concerns over the security risks associated with increasingly capable AI agents.
OpenAI has also placed greater emphasis on alignment and safeguards as its models become capable of performing more complex tasks.
Read more: OpenAI to end Cursor model access after SpaceX acquisition
Astra targets software engineering
OpenAI also described Astra as its strongest model for software engineering.
According to benchmarking results shared by the company, Astra outperformed existing models, including OpenAI’s Sol and Anthropic’s Fable, on tasks involving software bugs, terminal-based activities and questions about codebases.
The results suggest that OpenAI is positioning Astra not only as a general-purpose chatbot but also as a tool for developers and software engineers.
Opaque recurrence raises concerns
Astra’s capabilities have also prompted questions about a reasoning technique known as opaque recurrence.
The technique can make it more difficult for researchers to observe and audit how a model reaches a decision.
OpenAI has downplayed the extent to which Astra relies on opaque recurrence. However, the company acknowledged that monitoring AI reasoning becomes increasingly difficult as models grow more capable.
OpenAI Chief Scientist Jakub Pachocki said monitoring model reasoning remains an important part of AI oversight.
He noted that more capable models may sometimes complete difficult tasks with fewer language tokens, or without using language tokens at all. This could make their internal processes harder for researchers to monitor.
The issue has renewed debate over how researchers can maintain effective oversight as AI systems become more advanced.
Does Astra represent AGI?
Astra’s capabilities have also raised questions about whether OpenAI considers the model to represent artificial general intelligence, commonly known as AGI.
Brockman said AGI is no longer tied to a contractual trigger following changes to OpenAI’s agreement with Microsoft.
The previous agreement included a provision that could have affected the partnership if OpenAI achieved AGI. That requirement no longer exists, according to Brockman.
He said AGI has instead become more of a mission or “spiritual” concept, leaving individuals to decide whether a particular AI system meets their definition.
Brockman nevertheless said that, in his personal view, OpenAI has reached that point with Astra.
The comments add another layer to the debate over how AI companies should define AGI as their systems become capable of performing increasingly complex tasks.
Read more: realme report uncovers Pakistani GenZ’s responsibilities
Key points:
- OpenAI has launched Astra, which it describes as its most capable and intelligent AI model so far.
- The model improves computer and browser use, with a focus on speed, accuracy and safety.
- Astra is initially available to customers using OpenAI’s Daybreak cybersecurity programme.
- OpenAI plans to expand access to Pro, Plus, Enterprise and Business users before making Astra available through its API.
