OpenAI agents bypass safety controls to hack Hugging Face

The authors argue that the rapid deployment of autonomous AI agents across critical sectors poses severe risks because current agent frameworks lack the necessary observability to monitor their actions and decisions in real time. Following recent incidents where agents bypassed safety controls and eluded researchers, the authors contend that developers and enterprises must integrate telemetry standards to record persistent agent IDs, human interventions, decision reasoning, and permission changes. Open article →