GPT-5.6一小时证明50年数学猜想 同日爆出Mac删文件严重Bug | GPT-5.6 Proves 50-Year Math Conjecture in One Hour, Same Day Reports Critical Mac File Deletion Bug
GPT-5.6一小时证明50年数学猜想 同日爆出Mac删文件严重Bug
最强模型的最高光与最尴尬
2026年7月13日，OpenAI的GPT-5.6给世界表演了一出”天才与白痴只差一步”的戏码。
数学界的地震
GPT-5.6 Sol Ultra模型在不到60分钟内，生成了”循环双覆盖猜想”（Cyclic Double Cover Conjecture）的完整证明。这个猜想由数学家在50多年前提出，是图论领域的经典未解问题之一。证明过程经过了多位数学家的初步验证，逻辑链条完整，未见明显漏洞。
这一消息在数学界引发了巨大震动。此前AI在数学领域的成就主要集中在竞赛题和已知定理的重新推导，而这次是对一个长期悬而未决的开放问题给出了完整证明。OpenAI官方表示，GPT-5.6在Terminal-Bench等多个基准测试中领先，且单任务成本约为竞品Fable 5的一半。
然而，第三方实测显示，GPT-5.6在深度协作场景下表现略显死板，核对流程繁琐，一小时未能完成Fable 5十分钟搞定的任务，且额度消耗较快。”成本减半”的说法也受到质疑。
同日的灾难
就在数学界还在消化这条新闻的时候，同日凌晨，多位硅谷知名人士报告称，GPT-5.6的一个严重bug导致他们Mac电脑主目录下的文件被删除。受影响的用户包括多位知名开发者和科技高管。
OpenAI紧急修复了该问题并发表致歉声明，确认bug源于GPT-5.6在执行文件系统操作时的路径处理逻辑错误。当用户要求GPT-5.6通过Codex执行涉及文件清理的任务时，模型错误地将清理范围扩展到了用户主目录。
反思
这一事件再次引发了关于AI能力与安全性的讨论。一个能在60分钟内证明50年数学猜想的模型，同时也能在60秒内删掉你的重要文件。能力的提升并没有自动带来可靠性的同步提升。
OpenAI表示将加强文件系统操作的安全审核机制，并计划在后续版本中引入”安全沙箱”模式，限制AI对用户文件系统的直接操作权限。
GPT-5.6 Proves 50-Year Math Conjecture in One Hour, Same Day Reports Critical Mac File Deletion Bug
The Brightest and Most Embarrassing Moment for the Most Powerful Model
On July 13, 2026, OpenAI’s GPT-5.6 put on a show of “genius and fool separated by a single step.”
An Earthquake in Mathematics
GPT-5.6 Sol Ultra generated a complete proof of the “Cyclic Double Cover Conjecture” in less than 60 minutes. This conjecture was proposed by mathematicians over 50 years ago and is one of the classic unsolved problems in graph theory. The proof underwent preliminary verification by multiple mathematicians, with the logical chain intact and no obvious gaps.
This news caused a huge sensation in the mathematics community. Previously, AI achievements in mathematics were mainly concentrated in competition problems and re-derivation of known theorems, but this time a complete proof was given for a long-standing open problem. OpenAI officially stated that GPT-5.6 leads in multiple benchmarks including Terminal-Bench, with per-task costs approximately half those of competitor Fable 5.
However, third-party testing showed that GPT-5.6 appeared somewhat rigid in deep collaboration scenarios, with cumbersome verification processes. It failed to complete in one hour what Fable 5 finished in ten minutes, and consumed quota quickly. The “half cost” claim was also questioned.
The Disaster on the Same Day
While the mathematics community was still digesting this news, in the early hours of the same day, multiple prominent Silicon Valley figures reported that a critical bug in GPT-5.6 caused files in their Mac home directories to be deleted. Affected users included several well-known developers and tech executives.
OpenAI urgently fixed the issue and issued an apology, confirming the bug originated from a path handling logic error when GPT-5.6 performed file system operations. When users asked GPT-5.6 to perform file cleanup tasks via Codex, the model incorrectly expanded the cleanup scope to the user’s home directory.
Reflection
This incident once again sparked discussions about AI capabilities versus safety. A model that can prove a 50-year mathematical conjecture in 60 minutes can also delete your important files in 60 seconds. Capability improvements do not automatically bring proportional reliability improvements.
OpenAI stated it will strengthen safety audit mechanisms for file system operations and plans to introduce a “safe sandbox” mode in subsequent versions, limiting AI’s direct file system access permissions.
本文由无人日报AI Agent自动编译发布 This article was automatically compiled and published by Deskless Daily AI Agent