Google's WebMCP turns websites into agent tool surfaces

An AI Engineer talk walks through Google's WebMCP proposal, which lets a website declare its own controls as explicit tools an agent can call from inside the page. It's a different model from server-side MCP, and it raises immediate questions for anyone building agent-facing web apps. Treat it as a standards watch item: the demo runs in Chrome Canary behind preview flags, not in shipping browsers.

Kimi K2.7-Code ships as an open coding model

Hugging Face

Moonshot's new open coding model leads with token-efficiency claims and comparisons to paid options like Claude. The benchmark numbers come from launch and community material, so read them as vendor claims until someone reproduces them.

Kimi 2.7-Code release thread, with a benchmark figure

Reddit

The community sighting puts a specific number on the release, a claimed +31.5% on MLS Bench Lite. Same caveat applies: it's a launch claim, not an independent result.

Huawei previews openPangu 2.0, open weights on June 30

Reddit

Huawei put up a large sparse model with a 512K context window and said the weights open at the end of the month. For now it's a preview with metrics, not a download.

Nvidia's China-facing CPU availability

Techmeme

A market-and-supply-chain read on what hardware is said to be available to China. Separate the reported availability from confirmed shipments before drawing a trend line.

China's position in the robot supply chain

Techmeme

The physical-world side of the same story: who supplies the parts that go into robots. A separate signal from the chip and model items, not one grand strategy.

A judge removes lawyers from a case over AI-hallucinated citations

Forbes

A concrete professional consequence: fabricated citations got counsel pulled off the case. The clearest sign yet that high-stakes fields are setting standards faster than the tools verify themselves.

An auditable architecture for verifying biomedical manuscripts

arXiv

A research direction pointed at the same gap from the other side: build verification gates into large language model output for high-stakes writing. A preprint signal, not a fix for the courtroom problem.

Compliance-by-construction for agents in regulated industries

arXiv

One preprint makes the case that finance and health agents should be built so compliance is structural rather than checked after the fact. Research signal for teams in regulated environments.

Structural safety gaps in major agent frameworks

arXiv

A preprint arguing that frameworks like LangChain and AutoGPT don't constitute a safety case on their own. Mark it a research signal, not operational consensus.

A certified defense against memory poisoning in persistent agents

arXiv

Targets a real failure path for agents that keep state: poisoned memory. Proposes a certified defense the authors call SMSR.

ToolSense: diagnosing what an LLM actually knows about its tools

arXiv

An open framework for probing what a model knows about its tools past standard benchmarks, aimed at the reliability gaps that show up once an agent has to use real ones.

About 400 AUR packages compromised with an infostealer and rootkit

Hacker News

A reminder that the dev-tooling supply chain takes hits that have nothing to do with AI. Relevant if you maintain Arch machines or pull from community package repos.

A reported PeopleSoft flaw and associated breach claims

Techmeme

The enterprise-application side of the same week: a reported vulnerability with breach claims attached. Separate problem from the AUR packaging issue, same maintenance burden.