Agentic AI doesn't usually back up before acting (www.unite.ai)

A recent paper from a US academic collaboration highlights a concerning trend in the behavior of advanced AI models: a lack of backup protocols before executing actions that could result in significant data loss. The study found that leading models like GPT-5.6-Terra and Gemini 3.7-Flash backed up critical data in only 4.4% and 2.4% of cases, respectively. This negligence has real-world consequences, as demonstrated by incidents where AI agents inadvertently deleted crucial infrastructure and data, emphasizing the urgent need for more robust backup awareness in AI systems. To tackle this issue, the researchers introduced "BackUpper," a supportive framework that operates in conjunction with AI agents. This tool ensures that backups are made before risky actions based on a set of guidelines, including calling upon smaller AI models for complex scenarios. In tests, BackUpper achieved an impressive 92.2% similarity in data restoration while operating at a minimal cost of $0.014 per task. The study advocates for enhancing backup awareness through training and integrated support to mitigate costly and irreversible data losses, ultimately paving the way for safer and more reliable AI deployments in real-world applications.