Aleksander Madry Plans OpenAI Return to Focus on AI Safety

Aleksander Madry plans to return to OpenAI and work with Chief Scientist Jakub Pachocki on AI safety. Madry previously led a team that studied risks, and he is returning as the company deals with the cybersecurity abilities of advanced AI models.

Madry announced his return on Oct. 19, several months after leaving OpenAI. He said safety and security risks had become a major priority as AI development entered what he called a “different regime.” Pachocki asked him to study how quickly AI was advancing, what effects this progress could have, and how OpenAI should use this information to plan its strategy.

Madry previously led Preparedness, an OpenAI team created in 2023 to study the risks of advanced AI models. The team's work covers cybersecurity, biological threats, and the risk that AI systems could act more independently.

In September, OpenAI rated Astra's cybersecurity abilities as Critical under its Preparedness Framework. Its tests found that, with the right tools and access, Astra could find security flaws that were not known before and create ways to exploit them across many well-protected systems, without a person guiding every step. OpenAI delayed parts of Astra's development and release while it improved and tested protections against misuse and actions without permission.

This cybersecurity assessment gives Madry's return a clear focus. The main concern is whether a model can find and exploit weaknesses across many systems with little human guidance. This is one of the risks that the Preparedness team is responsible for studying.

Other safety concerns involve how AI systems act when they can use tools and interact with the outside world. OpenAI and Anthropic previously reported incidents during AI agent testing in which AI agents got around safety rules or reached real outside systems. Pachocki has called for closer monitoring. He has also said that the speed of AI model development should depend on how confident the company is in its safety measures.