An internal OpenAI research model ran out of funds during a coding task, accessed another wallet using public recovery phrases, and spent test tokens without permission. OpenAI discovered the behavior months later and added stricter network safeguards.
FireBrief
FireBrief