Moonshot AI has launched Kimi K2.7 Code, a massive open-weights model specifically designed for programming tasks. This new model features one trillion parameters and utilizes a Mixture-of-Experts architecture to optimize performance. It is currently available for developers looking for high-capacity alternatives to proprietary coding assistants.

Source

The Hades malware campaign has introduced a novel evasion technique that uses prompt injection to trick AI-based security scanners. Malicious JavaScript files now include code comments that instruct AI bots to generate instructions for creating biological and nuclear weapons. These prompts are designed to trigger the AI's safety failsafes, causing the scanner to halt the process before it reaches the actual malicious payload.

Source

Google has released the Google Colab Command-Line Interface to bridge local terminal environments with remote Colab runtimes. This tool allows for the provisioning of high-performance compute resources, such as GPUs and TPUs, directly from a local machine. It facilitates the execution of local Python scripts on remote hardware and the subsequent retrieval of generated artifacts or logs.

Source

Perplexity has introduced a new architecture called Search as Code that enables AI models to generate their own search workflows using Python. Instead of relying on static search APIs that return fixed lists of results, models can now write custom scripts to handle complex research tasks. This approach utilizes an Agentic Search SDK that provides modular functions for retrieving, filtering, and reranking data within a secure sandbox environment.

Source

A critical vulnerability in the Hugging Face Transformers library, tracked as CVE-2026-4372, allows attackers to execute arbitrary code through specially crafted model configuration files. The flaw bypasses the standard security setting intended to block untrusted remote code by exploiting how the library processes internal attributes during model loading. This vulnerability specifically targets the from_pretrained() function, which is a standard method used to initialize pre-trained models from the Hugging Face Hub.

Source