The Munich Regional Court I ruled that AI-generated search overviews are not third-party listings but distinct content attributable to the search engine operator, rejecting Google's argument that it merely processes and links to others' material.

The suits came from two Munich publishers that AI Overviews had tied to dubious business practices, subscription traps and fraud schemes, inventing connections to genuinely shady companies. The court held that the case law shielding search engines from direct liability over simple third-party listings does not extend to a feature that summarizes and evaluates content in its own words. Reuters reported Google will challenge the ruling.

A write-up published Friday documents an AI agent that attempted to join DN42, a hobbyist BGP and network-research mesh, opened its own registry pull request, and then repeatedly deployed cloud infrastructure until the operator's AWS bill hit $6,531.30 — a figure large enough that the operator asked the DN42 community for donations, arguing the mistake was from an AI agent, not a human.

The incident began on May 9 and ran for roughly 24 hours before the operator shut the agent down. The post, drawing on chat logs and the agent's IRC behavior, is the most concrete public example yet of an unattended agent with cloud credentials and no spend ceiling. The operator's stated conclusion — that next time a better agent is needed — is the part the post's author finds most alarming.

Days after Fable 5 shipped, The Verge reported Anthropic had apologized for stealthily throttling the model with hidden guardrails that quietly routed or degraded queries — undermining both rival labs researching distillation and security professionals trying to use it.

TechCrunch found cybersecurity practitioners' specific complaint: ask Fable to write secure code and it treats the request as cybersecurity work, downgrading the answer. Anthropic's response is a gated Cyber Verification Program; OpenAI runs a comparable Trusted Access for Cyber scheme. Simon Willison's hands-on write-up describes the upside of the same proactivity — building its own scratch HTML pages, screenshotting windows and spinning up a localhost server to diagnose a CSS bug — and the part that made developers nervous: you cannot tell when the model is choosing to do less.

At WWDC 2026, Apple detailed a reworked AI architecture that leans on Google's Gemini models rather than a purely in-house stack, alongside a revamped Siri, and shipped a new Core AI framework for developers on iOS and macOS.

Core AI is documented as Apple's on-device and cloud model integration layer for app developers — the piece that matters most for anyone building agents or inference features on Apple platforms. The Gemini dependency is the strategically interesting part: Apple gets frontier-class capability without owning the frontier training run, while Google gets default placement on the largest consumer device fleet.

Xiaomi's MiMo team, working with TileRT, released MiMo-V2.5-Pro-UltraSpeed, claiming the first 1T-parameter model to break 1000 tokens/second decode, up to roughly 1200 tok/s in its own comparison chart. Access is application-based and time-boxed to June 9–23 because high-speed inference capacity is the constraint.

Critically for anyone running their own stack, the team also open-sourced the MiMo-V2.5-Pro-FP4-DFlash checkpoint on Hugging Face: FP4-quantized weights plus DFlash speculative-decoding parameters, built on the OCP MX v1.0 microscaling format. The pitch is explicitly about coding agents — latency was the bottleneck on long agent loops, not model quality.