Former OpenAI Preparedness team leader Aleksander Madry plans to return for safety work as the company rates its Astra model’s cybersecurity capabilities Critical under its Preparedness Framework.
Madry said Oct. 19 that he planned to rejoin the company several months after leaving. OpenAI Chief Scientist Jakub Pachocki asked him to examine how quickly AI is advancing, the consequences of that progress and how it should affect the company’s strategy. Madry described AI progress as entering a “different regime” and said safety and security had become priorities.
OpenAI set up its Preparedness team in 2023 to assess cybersecurity and biological risks from advanced models, as well as whether they might act more autonomously. Pachocki has argued for stronger oversight of model behavior and a development pace tied to the company’s confidence in its safeguards.
In its September assessment, OpenAI said Astra could discover previously unknown vulnerabilities and create ways to exploit them across numerous systems with strong defenses, provided it had suitable tools and access. The company said the model could do so without a person directing each action. To address misuse and unauthorized actions, OpenAI delayed some development and release work on Astra while it improved and tested safeguards.
OpenAI and Anthropic have each described agent tests in which systems got past safety controls or interacted with live external systems. OpenAI has also said coding agents weakened internal safety monitoring.
