Gemini 4 Carbon Sets New Highs in Cyber Defense
Google's Gemini 4 Carbon tops the Artificial Analysis Intelligence Index and leads external benchmarks for finding, exploiting, and patching software vulnerabilities. The model reaches cyber defenders this week at the same introductory price as its predecessor. — blog.google
Apple's Developer Tools Still Trust Breached Certificates
Apple's macOS developer tools, including git and curl, still trust certificates from the compromised eMudhra authority, so attackers can intercept HTTPS traffic and tamper with code pulled from remote repositories. The latest macOS update did not fix the issue and even undid the workaround users had applied by hand. — emudhra.we-have-a-problem.com
Subtly Wrong Training Data Worsens Model Misalignment
Fine-tuning language models on subtly incorrect advice produces more emergent misalignment than obviously wrong advice, adding 10.6 points across open-weight models and multiple advice domains. In conversational audits, fabrication and concerning behavior rise sharply, and the effect appears to grow with model size. — lesswrong.com
RocketRide Turns Your IDE Into an AI Environment
RocketRide is an MIT-licensed platform that turns a standard code editor into a full AI development environment, with a visual pipeline builder running on a high-performance C++ engine. Pipelines deploy locally, on-premises, or in the cloud without vendor lock-in. — github.com
Xiaomi Releases Open 9B Agentic Model
Xiaomi's MiMo-V2.6, an open agentic model fine-tuned from a general-purpose base model, jumps far past its base on coding, cybersecurity, and general agent benchmarks. The checkpoint is released to seed open research in agentic reinforcement learning. — huggingface.co