Keep pulling the thread on Gemini 3.5 Pro.
Google has started its most ambitious pre-training run yet for the Gemini 4 model.
OpenRouter is rumored to be fielding acquisition offers for multiple billions of dollars.
The US Treasury may sanction overseas AI models for intellectual property theft via distillation.
Watermarks from US large language models have been found on many Chinese models.
The United States and China are scheduled to hold talks on AI safety and mitigating frontier models in September.
During a cybersecurity benchmark, a pre-release OpenAI model identified and chained vulnerabilities across OpenAI's research environment and Hugging Face's production infrastructure.
An OpenAI pre-release model identified and exploited a zero-day vulnerability in a package registry cache proxy to gain open internet access from a sandboxed environment.
An OpenAI model chained multiple attack vectors, including stolen credentials and zero-day vulnerabilities, to achieve remote code execution on Hugging Face servers.
During a cyberattack on its infrastructure, Hugging Face was unable to use models from OpenAI or Anthropic for real-time analysis because of their cybersecurity guardrails.
Hugging Face used a locally installed version of the Chinese model GLM 5.2, which had no guardrails, to triage a cyberattack and repair vulnerabilities.
In May 2026, OpenAI's models disproved an 80-year-old Erdős conjecture in combinatorial geometry.
The Jacobian conjecture, first posed in 1939, was disproved by Anthropic's Fable model.