Technology News

A new study from Google Research and Technion says frontier models often already encode facts they fail to state, shifting attention from missing knowledge to recall.

OpenAI says Astra is its first model to hit its “critical” cyber threshold, prompting a restricted launch and early access for security partners.

OpenAI says it paused parts of its unreleased Astra model work to strengthen cybersecurity safeguards after an earlier model escaped its limits and hacked Hugging Face.

A real Azure OpenAI email assistant passed testing but still returned SharePoint files the requesting user could not access, highlighting a permissions gap in retrieval pipelines.

Perplexity’s new hybrid AI system splits work between cloud and local models on Apple silicon Macs, keeping sensitive data on the device.

Google has raised the Google TV Streamer’s price to $149, putting the 4K box in line with a broader wave of streaming-device increases from Apple and Amazon.

AI coding agents are changing software work. As they generate more first-pass code, engineers are increasingly needed to define constraints, tests, and boundaries that keep systems reliable.

Nvidia is investing $3.5 billion in MediaTek to help custom AI chips plug into Nvidia-based data centers, even as Big Tech builds more of its own silicon.

James Hall argues that more AI workloads should move from cloud servers to local devices, using browser tools like WebGPU, Transformers.js, and DuckDB to improve privacy and performance.

As AI agents take on more autonomous work, enterprises are being pushed to rethink security. The key issue is not just identity, but whether an agent’s actions stay trustworthy after login.