Radware disclosed vulnerabilities letting an attacker exfiltrate data from connected services — Gmail, Outlook, Drive, GitHub and others — without the victim clicking anything, and persist the bypass logic by writing it into ChatGPT's long-term memory.

OpenAI had already tried to break the attack by forbidding connectors and memory in the same session and blocking memory-sourced URL opens; Radware says the guardrails remain bypassable.

In his CES keynote Jensen Huang introduced Nvidia's next platform — six new chips fabricated by TSMC — and said it is now in full production, with the focus on the Vera CPU plus Rubin GPU pairing and the rack-scale design changes that come with it.

Coverage out of Taipei framed it as the company moving from sampling to volume on the hardware that will underpin the next wave of training and inference clusters.

xAI completed an upsized Series E, raising $20B against an initial $15B target, with reports putting the post-money valuation around $230B and Nvidia and Cisco among the backers.

The company framed the raise around model development and its Colossus data-center buildout. It landed while xAI was under scrutiny over child-safety issues around Grok.

Nvidia released open models, datasets and tools across four families: Alpamayo for autonomous vehicles, Nemotron for agentic AI, Cosmos for physical AI, GR00T for robotics and Clara for biomedical. The centerpiece is Alpamayo 1, a 10B chain-of-thought vision-language-action model that reasons about a driving maneuver before emitting a trajectory.

Alongside the weights Nvidia contributed roughly 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures and about 100TB of vehicle sensor data, plus an AlpaSim simulator and 1,700+ hours of driving data.

TII released Falcon-H1-Arabic at 3B, 7B and 34B, built on the Falcon-H1 hybrid architecture that runs Mamba state-space layers and Transformer attention in parallel inside every block and fuses their representations before the output projection.

TII claims the family beats same-size and sometimes larger models on Arabic benchmarks, with the hybrid design giving substantially better long-context memory behavior and faster inference than a pure Transformer of comparable quality.