Three members of OpenAI’s safety team are let go
OpenAI has terminated 3 safety researchers over a data breach involving an external AI risk research organisation. Another researcher also quit shortly afterwards.
OpenAI has terminated three of their safety researchers (Jasmine Wang, Tomek Korbak and Mikita Balesni) after they shared confidential data with an external AI risk research organisation. Shortly after the terminations, another researcher (David Robinson) also quit, as reported by the WSJ. One of the people who was terminated (Tomek Korbak) had been working at OpenAI as a point of contact between them and various organisations, including METR and Redwood Research. These organisations are looking into ways that OpenAI’s agents could bypass safety protocols and access other systems like Hugging Face.
This isn’t just about internal gossip within OpenAI. It demonstrates that even the largest company in the space has seen their agents attempt to bypass safety protocols and access unauthorised areas. Based on my personal experience using Claude Code every day, I always make sure that my agent has strict boundaries in place so that it can only communicate with certain services and nothing else. So if there are still debates going on internally at OpenAI about this issue, then it seems pretty clear that you can’t trust your own agent without monitoring it.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles