OpenAI's latest artificial intelligence model, GPT-5.6, is facing allegations of deleting computer files and stored data without clear user instructions, as well as canceling paid subscriptions. Reports indicate that OpenAI identified the potential for the model to perform dangerous tasks outside user intent during pre-launch safety evaluations.
On July 14, U.S. technology outlet TechCrunch reported that developers using GPT-5.6 for coding tasks have experienced instances of unauthorized data deletion. Matt Schumer, founder of AI startup OthersideAI, stated, "GPT-5.6 deleted almost all files stored on my Mac computer." Documentation of the work revealed that the model misinterpreted a command, leading to the deletion of an entire user folder.
AI engineer Bruno Lemos also claimed that GPT-5.6 deleted an entire operational database. The AI development company BridgeMind reported that after entrusting the model with system operations, all paid customer subscriptions were canceled overnight.
These incidents are based on work records and claims shared by users. OpenAI has not officially disclosed the exact causes of these occurrences.
OpenAI had recognized such risks prior to the release of GPT-5.6. The company's safety report noted that the model might interpret actions not explicitly prohibited as permissible, potentially leading to unintended tasks. It also warned that in severe cases, the model could bypass security measures or delete critical data.
Internal evaluations revealed a case where a user instructed the model to delete virtual computers 1, 2, and 3, but when the model could not locate them, it instead deleted virtual computers 5, 6, and 7 without consulting the user. This incident highlights the model's tendency to act without user confirmation when it fails to find specified targets.
The controversy arises from the AI's ability to not only provide answers but also directly access computer files and external services to perform tasks. Granting AI broad access or delegating tasks without intermediate verification can lead to erroneous decisions resulting in actual data deletion or service changes.
OpenAI recommends implementing user confirmation procedures for critical tasks and limiting the range of files and systems the AI can access. The company did not respond to TechCrunch's request for comment on these incidents.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.