Tom's Hardware – AI

Deepseek 4.1 Flash 1017 drives cost-per-task further down still — open-weight model is fit for agentic workloads and exceedingly cheap to run

Several U.S. banks are blocking AI agents from making transactions on their platforms. They fear that those agents could go rogue and there's no way to tell if the moves they make are legitimate or not.

Claude and Grok are among the AI agents that are being used to automate attacks on Latin American institutions, including banks and payment apps, by a group calling themselves CyberCartel.

Wolves, who used to live in the forests recently cleared for Google data center development, are now roaming into Finnish villages nearby.

OpenAI says researchers using GPT-5.6 Sol uncovered two critical PHP vulnerabilities with severity scores of 9.2 and demonstrated working exploits, after which the company coordinated fixes.

Google’s TPUs now serve AI in production at Meta, through new open-source engines that run PyTorch models natively on the chips.

The U.S. Army is developing a program to deploy autonomous AI agents that can navigate digital and network infrastructures, identify targets, and “neutralize” them.

The U.S. government is helping fund projects to bring nuclear reactors back online in an attempt to provide more energy capacity to AI data centers.

Google DeepMind researchers say four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs with no human help.

A Chinese hacker reportedly used a fleet of AI agents to break into more than 100 company websites, including farm equipment maker Deers and appliance outfit SharkNinja.

Arm claims its AGI CPU delivers nearly 2X the modeled rack performance of AMD’s EPYC 9755 across seven agentic AI workloads, using a new evaluation methodology developed with Fujitsu, Meta, and OpenAI

Anthropic's proposal requires evaluators to report any imminent risk to the Department of Commerce within 72 hours, after which the Department must present the evidence to a federal judge and await an order.

Researchers have developed WiFireNet, an AI system that detects concealed handguns from Wi-Fi channel-state information, achieving 72.5% average precision in controlled indoor tests without cameras or metal detectors.

HUMXN wants to pay for maintenance and repair services for homes in Minneapolis, Chicago, and Miami to gather data to train AI models and robots.

GPT5.6-Sol didn't have funds to deploy smart contracts, so it looked for exposed accounts with enough money to pay for the gas fees.

ASML says customers are already placing orders for advanced chipmaking tools years in advance, with AI-driven semiconductor demand remaining strong enough to strain capacity despite planned expansions

The Challenge Sweden's season 2 edit was deleted by an AI agent put on the production server. The program director blames "the human factor" instead of AI.

Apple researchers describe Minions, an internal system that runs AI inference jobs on idle Apple Silicon Macs when GPUs are scarce.

Anthropic’s paper establishes conceptual model capabilities, but doesn’t quite demonstrate a real threat until models improve further.

Last week, OpenAI shared over 400 results for longstanding mathematical formulas that have troubled humans for decades; now researchers say it could take years for humans to validate and prove all the AI-driven work