Microsoft treats AI agents as insider threats with Guardian Agents
New Defender and Guardian Agent features monitor and restrain workplace AI agents, as Nadella argues intelligence supply must be separated from authority over it.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse Crunchy Tech’s latest ai briefs, grouped by date.
PUBLISHED
PUBLISHED
Amazon's new tool lets PyTorch apps run on its Trainium AI chip with a one-line change, as AWS also pushes to make PyTorch less Nvidia-centred.
The 120-billion-parameter Super VL targets enterprise agents, while two smaller Bolt models are built to run locally on RTX Spark and DGX Spark machines.
AI agents can now remove bloatware, sponsored ads and tracking from smart TVs over Wi-Fi, but the hacks remain works in progress with real risks.
PUBLISHED
PUBLISHED
Yelp, Delta, JPMorgan Chase and others are restricting AI agents, while anti-bot systems and legal fights with Amazon complicate the agent boom.
Bee Frontier and Bee Eclipse turn conversations into summaries, to-dos and reminders, and now link to Alexa+ and Amazon for shopping recommendations.
PUBLISHED
Stonking Stingray pairs GNOME 51 and Linux 7.3 with an opt-in, locally run speech-to-text feature that Canonical plans to make standard in future releases.
New open models from Reflection, Mistral and Aleph Alpha aim to win security-conscious users, but experts say Chinese systems remain ahead on performance and price.
Nick Cammarata, who helped shape the superposition hypothesis, now says AI models are far easier to interpret than the field believed. Not everyone is convinced.
The new lab's first model is a sparse mixture-of-experts system for coding and agentic work, released under Apache 2.0 with a 1M token context window.
Torvalds used an AI model to tackle a long-standing cache contention problem, claiming a 23x gain in a synthetic benchmark.
PUBLISHED
Black Forest Labs' 5.6-billion-parameter model handles text-to-image, editing and reference-based generation, with an INT8 build running on consumer graphics cards.
On the TheoPhysEval benchmark, the AI agent matched leading specialists on unpublished theoretical physics calculations, at roughly $2.40 per problem.
A 120-billion-parameter open vision-language model for enterprise agents arrives with two compact Bolt variants for local PCs and edge devices.
A "Made with AI" tag is rolling out in the YouTube Music app, with the main YouTube app to follow.
PUBLISHED
Epic's rebuilt Epic Games Launcher enters preview on Windows, Mac and Linux, built with AI and claiming faster startup than Steam.
Probabilistic Bias Correction uses machine learning to fix systematic errors in existing weather models, improving accuracy two to six weeks ahead and running on modest hardware.
A toggle due before the end of the year will remove AI tracks, albums and artists from search, as the company says machine-made uploads now exceed half of its daily total.
An internal system called Minions redistributes stalled AI jobs to spare Apple Silicon Macs, cutting some waits from hours to minutes.
A privacy-preserving Gemma model fine-tuned on local scam conversations will warn Singaporeans in real time, with rollout phased alongside GovTech.
PUBLISHED
Shoppers can connect their own AI agents to Shop and buy via Shop Pay, no partnership required, with approval required on every purchase.
An open letter from leading AI researchers urges full network isolation, binding containment standards and public reporting of breakout incidents.
Google's long-awaited model claims frontier-level performance at a fraction of rival pricing, though it is initially limited to Pro and Ultra subscribers.