METR and Redwood Research found monitoring gaps in OpenAI systems
Between July and August 2026, outside safety groups METR and Redwood Research were invited to poke around OpenAI's internal systems, and they found some real weaknesses.
Their team discovered new ways to break the monitoring setup, potentially letting certain actions slip by unnoticed.
The testers had in-person access to OpenAI's code and staff, but everything was tightly controlled on company devices.
OpenAI patches safeguards, finalizes assessor contracts
OpenAI moved fast once the issues were flagged, patching up its safeguards. The exercise was part of a broader effort, with Sam Altman, CEO of OpenAI, promising more outside oversight.
METR and Redwood handed over a detailed report, which OpenAI says will help make its systems safer going forward.
They're also actively finalizing contracts with third-party safety assessors, showing they're serious about tightening up security and working with third parties to keep AI development in check.