Google releases HEIR, an open-source compiler for running AI on encrypted data
Google released HEIR (Homomorphic Encryption Intermediate Representation), an open-source compiler toolchain that converts pre-trained AI models to run inference on encrypted inputs, so the server processing a request never sees the underlying data. Google is aiming for a "one-click" experience so non-cryptographers can add encrypted inference to production apps, and has partnered with hardware accelerator makers (Belfort Labs, Niobium, Cornami, Optalysys) plus Georgia Tech, Carnegie Mellon, and UC Santa Barbara. Demonstrated uses include fraud detection, network intrusion detection, and voice-assistant hotword recognition.
Why it matters: Homomorphic encryption has existed for years but stayed a research curiosity because using it required a team of cryptographers; a usable compiler is what actually determines whether "private AI inference" becomes a real deployment option or stays a talking point.
Claude Code makes "Auto Mode" the default permission setting
Anthropic made Auto Mode the default for Claude Code's Pro, Max, and Team users, replacing repeated manual approval prompts with a classifier that flags irreversible, destructive, or out-of-bounds tool calls for review while letting routine actions through automatically. Anthropic says Team and Enterprise users are shipping about 25% more pull requests since adopting it, and stopped charging extra tokens for the classifier last week. Anthropic still recommends human review before production changes.
Why it matters: Cutting approval friction while keeping a safety classifier in the loop is the shape most agentic coding tools will need to converge on if they want both speed and a defensible "we didn't just let it do anything" story.
Meta patents AI smartglasses that identify people and auto-generate highlight reels
A Meta patent surfaced by 404 Media describes AI glasses that use facial recognition to identify people in the wearer's field of view, capture a clip whenever a recognized person does something (picks up an object, moves), and stitch the footage into an automated highlight reel — for example, of a dinner party. The system can also read facial expressions and use "user relationship data" to personalize what gets included. Meta has not confirmed the feature will ship.
Why it matters: A patent isn't a product, but it's a specific, working description of what "AI memory glasses" would need to do technically, and facial recognition plus automated footage compilation is exactly the combination privacy regulators have flagged hardest in prior smart-glasses controversies.
OpenAI confirms ads are coming to ChatGPT's Free and Go tiers in Europe
OpenAI Ireland notified Free and Go plan users across the EEA and Switzerland that ads will begin appearing in ChatGPT later this month. Ads won't be personalized at first in the EEA/Switzerland specifically, using only conversation topic, general location, and device type as signals; Pro, Business, Enterprise, and Edu accounts stay ad-free. The move follows earlier ad rollouts in the UK, Japan, South Korea, Brazil, and Mexico.
Why it matters: Free-tier ChatGPT users are now the product in the same sense as any ad-supported platform, and the EEA rollout is the test case for whether OpenAI can run ads and still satisfy GDPR-level scrutiny.
AI Impact Hub