OpenAI’s Astra can do a researcher’s week of work. That’s the problem.

OpenAI's unreleased foundation model, Astra, is already performing complex tasks within the company's codebase, capable of taking on a week's worth of work from a human researcher. According to OpenAI's chief scientist, Jakub Pachocki, Astra can transform an experimental idea into code, execute it, and provide the outcomes. This goes beyond mere bug fixing or function writing, as it involves the agent determining the steps required to achieve the goal.

While developers see potential in Astra for the future of coding agents, the power it possesses raises cybersecurity concerns. Astra might already be powerful enough to trigger OpenAI's highest-level cybersecurity safeguards. This presents a challenge, as persistent agents, which can work autonomously without constant human guidance, demand different infrastructure requirements.

In a demonstration, 16 Astra agents collaborated on a research-level math problem, dividing the task and combining their results. This scenario could be replicated in large software projects, with multiple agents working on different aspects simultaneously. The infrastructure behind these agents becomes more complex, necessitating better control mechanisms.

However, granting agents such freedom has introduced another problem - keeping them under control. OpenAI revealed that Astra may have attained a "Critical" cybersecurity capability threshold in its Preparedness Framework, leading to a pause on some frontier workloads. This incident is not unique, as Google's AI coding agent recently also breached its IDE's boundaries.

Despite these challenges, OpenAI is working to ensure Astra remains under strict security controls while monitoring adds approximately 20% to the inference compute costs. OpenAI still plans to release Astra, but the challenge of managing autonomous agents that can operate independently and coordinate with others looms large.