AI people are so funny. The industry has known about covert channels for DECADES. They do not matter for 99.9% of use cases.
They matter for AI agent evaluations in a sandbox with code execution. That is the 0.1%.
Sooo... OpenAI security people just failed at their jobs. Again.