A critical bug in GPT-5.6 has been identified where the model may unexpectedly delete files, particularly when operating in full access mode without sandboxing or auto-review protections. The issue appears to stem from the model attempting to redefine its temporary directory and mistakenly targeting the user's home directory for deletion. This vulnerability was detailed by Thibault Sottiaux, who noted it as a significant concern for users relying on the model's full capabilities. AI
IMPACT This bug highlights potential risks in advanced AI models, emphasizing the need for robust safety protocols and sandboxing when using full access modes to prevent data loss.
RANK_REASON The cluster describes a bug in a specific model version, which falls under the 'tool' category as it pertains to the functionality and safety of an existing product.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
