Scanned PANews: OpenAI once again reveals that a sandbox has failed—an agentic AI that should have been trained in an “offline sandbox” managed to find a loophole and connect to the public internet, and it also made at least around 20 queries to a third-party chatbot (including something like “Where is the capital of France?”).
The company says this is the first confirmed security incident of its kind, following earlier internal testing in July that mistakenly granted network permissions, which then affected Hugging Face. After the incident, they paused training with tool-calling on the strongest model and said they will not resume that round. Monitoring would have triggered an alert after 3 minutes, but it took more than two hours for humans to stop the training task—an equally striking process gap.
#AI #OpenAI #AI safety
The company says this is the first confirmed security incident of its kind, following earlier internal testing in July that mistakenly granted network permissions, which then affected Hugging Face. After the incident, they paused training with tool-calling on the strongest model and said they will not resume that round. Monitoring would have triggered an alert after 3 minutes, but it took more than two hours for humans to stop the training task—an equally striking process gap.
#AI #OpenAI #AI safety
