Friday, November 6, 2026
CISA, the FBI, HHS, USDA and DOT say attackers are exploiting exposed temperature controllers and can force units into constant heating and defrost cycles.
The model did about 40% worse on safety tasks than on identical general tasks. OpenAI traced the problem to specific training data and says removing it fixed the behavior.
The bug, called "*sshole," affects Alpine Linux, OpenWrt, Buildroot and Yocto systems under certain authentication settings, and dates back to 2016.
Researchers say the flaw made the optional protection against prompt-injection data theft ineffective. In ChatGPT Classic, the attack needed no user interaction.
OpenAI says the model is better at finding vulnerabilities and building patches, and is trained to be more careful about user intent after a summer of sandbox escapes.
In independent tests, the model ranks second among Chinese models but trails Claude Opus 5.5 well behind on hard math and coding.
The agreement let Claude show live AP election results with attribution on midterm night. Anthropic had previously argued that scraping news was fair use.
The two companies responded to an election in which voters strongly opposed data center expansion.