OpenAI’s new flagship AI model unexpectedly deletes users’ local files and databases, as reported by TechCrunch.
Strangely enough, such glitches happen quite frequently, whether with massive models like GPT5.6 and Claude Opus 4.8 or smaller open-weight models.
Why does this occur? The issue lies in the very nature of these models: they are not “algorithmic” (consistently delivering repeatable, stable results like a calculator) but probabilistic. This means there is always a chance of an erroneous or unwanted response. And that undesirable “response” can manifest as file deletion. Ironically, humans share this same trait with modern AI. We are not calculators either; we might give different answers to the same question with a certain probability. Hence the outcome—”To err is human,” just as it is for large language models (LLMs).
Another interesting observation: for an LLM, the tasks of generating a reply to “hello, how are you?” and deleting all files on a device differ only in computational cost. Thus, an LLM can cause exactly as much harm as a user allows it to. That is precisely why developers conduct a special training phase for large language models called alignment, during which constraints are instilled. However, even alignment offers no guarantees, so restrictions are also built into the models’ software environment, noted Alexey Pekhterev, head of the special projects group at Positive Technologies’ ML department.
Therefore, before granting permissions for any actions, users should weigh the consequences. It is crucial to remember: if an agent is authorized to modify data, code, or settings, it may corrupt information, cause leaks, or introduce unsafe changes.
In such situations with AI models, the comforting aspect is that this is not some malicious intent or global conspiracy, but rather an error and a lack of proper LLM utilization. And if that is the case, the risk of AI-inflicted damage can be significantly reduced. This is facilitated by limiting permissions, vetting downloaded code, and overseeing execution. Currently, solutions are emerging for corporate clients that restrict and monitor agent behavior. Moreover, a specialized cybersecurity domain is taking shape today, focused on countering attacks on AI, noted Alexander Lebedev, senior AI developer at Innostage.
About the Author
