AI News, July 28: MCP Gets Its Biggest Rewrite Yet
Tuesday split cleanly into plumbing and consequences. The protocol that connects AI agents to everything else got rebuilt for scale, a European nonprofit demonstrated that the largest open model repository will undress people on request, and Korean chip stocks had their worst day since March on fears that have nothing to do with model quality.
The Big Story: The Protocol Layer Grows Up
The Model Context Protocol shipped its largest revision since Anthropic launched it twenty months ago, now under the Agentic AI Foundation, a directed fund of the Linux Foundation. Four changes matter. MCP completes its move to a fully stateless architecture, which kills the session-store requirement and lets servers sit behind ordinary load balancers. Issuer validation on the iss parameter becomes mandatory, closing the OAuth mix-up attack class — maintainers stressed this was preventive, with no known exploitation. There is now a formal twelve-month minimum deprecation policy, a number set after consulting Google, Microsoft and Amazon. And MCP Apps (server-rendered interactive UIs inside AI clients) and MCP Tasks (durable handles for long-running work, so a client can disconnect or crash and resume) graduate to official extensions.
Co-creator David Soria Parra put it plainly: “Some people jokingly call it a v2, and I think in spirit that’s accurate. It’s probably the biggest change we’ve ever made to the protocol.” Maintainer Den Delimarsky described what statelessness fixes: “Before, you needed to have a session store and manage session IDs — and if one of your compute pods went down, all of a sudden the requests would start failing.” AAIF executive director Mazin Gilbert framed the motive as pure scale: “I’ve come across companies who are deploying tens of thousands of agents, and you cannot do that without having to go in this direction.” The maintainers were unusually candid about trade-offs, too — wire payloads get larger, and out-of-band server logging is gone after they scraped all of GitHub and found usage amounted to “quite literally a handful of people.”
One caveat worth stating clearly, because it changes what you should do today. As of this writing the canonical artifacts have not caught up with the announcement: modelcontextprotocol.io/specification/latest still redirects to the 2025-11-25 revision, and the only 2026-07-28 tag in the spec repository is a release candidate published back in May. MCP names revisions by target date, so the version string has existed for two months. Read this as an announcement on embargo timing with the docs merge pending, and check the repo before you migrate anything.
Today’s Top Stories
Seven of Nine Hugging Face Image Models Will Undress Someone on Request
European nonprofit AI Forensics published a report finding that seven of the top nine image-editing models hosted on Hugging Face readily complied with simple prompts to undress women. WIRED’s version reports a six-word prompt was enough, and adds an analysis of a corpus of 1,000 image-editing prompts showing how people actually use these tools. The contrast the report draws is the uncomfortable part: Gemini and ChatGPT block these prompts, the Hugging Face-hosted models tested did not, and the repository is doing very little to prevent it. Three major outlets ran it within six hours, which makes this the day’s most-covered AI story — and it lands squarely in the middle of this week’s open-weights argument.
Korean Chip Stocks Crater on a Chinese Lithography Breakthrough
The KOSPI closed at 6,023.66, down 732.09 points or 10.84%, its worst session since March, dipping briefly to 5,992.91 intraday — the first sub-6,000 print since April 14. Samsung Electronics fell 13.39% and SK hynix 14.65%, and sell-side sidecars and circuit breakers triggered simultaneously across KOSPI and KOSDAQ for the third time this year. The trigger was specific and not about AI demand: a Shanghai state-owned enterprise began producing immersion DUV lithography equipment, with five units planned this year and twenty next. Given that SMIC has already run advanced processes on immersion DUV despite export controls, the technology gap could narrow faster than the memory market has priced in. Early Reuters coverage put the drop at “nearly 5%” — that was intraday, and worth flagging if you saw the smaller number.
Meta Puts a Gigawatt in El Paso and Keeps 20% of It
Meta announced a venture in which BlackRock-managed funds take 80% and Meta retains 20% of a Texas campus already under construction, designed for one gigawatt of compute. Total development cost is roughly $14 billion: Meta contributes land and in-progress construction worth about $2.3 billion, BlackRock puts in about $4.9 billion cash, and $12.5 billion in debt financing backs the structure. Meta receives a roughly $1 billion one-time distribution, signs a four-year initial lease with four extension options across a potential twenty years, and provides residual value guarantees with an aggregate threshold near $13 billion. Operations expected in 2028. The timing is not subtle — this closes days before Meta’s Q2 earnings, and it moves a gigawatt of capex off Meta’s balance sheet in the same week the market started punishing AI infrastructure debt.
Amazon Retires Most of the Nova Lineup
Amazon is deprecating Nova Premier, Nova Omni, Reel and Canvas, shifting resources into a new Frontier Model Research effort led by Pieter Abbeel. Nova 2 Lite, Nova 2 Sonic and Nova Forge survive, and a new flagship is expected at re:Invent later this year. It follows last week’s closure of Amazon’s San Francisco AGI Lab. Amazon spent two years trying to compete across every modality at once and has now conceded that was the wrong shape.
Two Papers Say the Harness Matters More Than the Model
Two results announced today should change how you read a coding-agent leaderboard. The Scaffold Effect fixes the model and varies the harness — Qwen 3.6 Plus and MiniMax M2.5 across Goose, OpenCode and OpenHands-SDK on a 50-task subset of Terminal-Bench Pro. Harness choice induces up to a 40× difference in tokens per solved task, while paired within-model pass-rate differences stay at 0 to 8 percentage points. Failure signatures replicate across models, meaning the biases belong to the scaffold, not the weights. Separately, StateAct argues a screenshot is a lossy render of program state and has its main agent work in code, handing off to a GUI subagent for only 28 of 108 tasks and 1.1% of steps. On OSWorld 2.0 that lifts Claude Opus 4.8 from 20.6% to 26.9% binary success at roughly 9× lower cost per task — and the ablation matters: a code-only variant with no GUI subagent reaches only 45.9% partial success, below the 54.8% screenshot baseline. The hybrid is doing the work.
Quick Hits
- Infrastructure: The Financial Times identified Nvidia as the anonymous tenant behind the lease on Hut 8’s one-gigawatt Beacon Point campus in Texas — $19.6 billion over a fifteen-year base term, rising to as much as $50.2 billion with renewals.
- Export controls: Taiwanese prosecutors detained an Nvidia employee on forgery and breach-of-trust allegations in a chip-smuggling probe now holding seven people, over roughly 50 Super Micro servers allegedly shipped to China on forged paperwork.
- Funding: Agent governance was the day’s clearest investment theme. Hush Security raised $30 million to enroll every agent in a registry with scoped just-in-time permissions, and Act Security disclosed $60 million across a seed and Series A for enforcing least privilege across humans, workloads and agents together.
- Liability: A law firm called Crosby says it will buy professional liability insurance covering its AI agents so they can work without a lawyer reviewing every output. No insurer or coverage amount named yet, so this is an intention rather than a policy — but insurability, not capability, is the real gate on autonomous professional work.
- Unit economics: LexisNexis CEO Sean Fitzpatrick said inference token costs are under 1% of RELX revenue, a useful counterweight to the assumption that inference will eat software margins.
- Sovereign compute: Firebird is building Armenia’s national AI capacity — $500 million and 6,000+ Blackwell GPUs in phase one, around $4 billion total, with Mistral AI signed on for government applications.
- Autonomy: Lyft and Baidu began robotaxi testing in London with dozens of vehicles and safety operators, targeting a public launch in 2027. Waymo has been testing there since April.
- Tooling: Agent sandboxing showed up twice on Show HN within one minute — Hotcell and Microsoft’s Quicksand, which sandboxes an agent without Docker or WSL. Between those, a browser layer, a backend provisioner and an MCP catalog, essentially every launch today was infrastructure for agents rather than an agent.
What This Means
The two halves of today are the same story told from opposite ends. At the protocol layer, MCP going stateless, mandating issuer validation and promising twelve months of deprecation notice is what a standard looks like when its users stop being demos and start being fleets — Gilbert’s “tens of thousands of agents” is the whole explanation. At the market layer, a lithography line in Shanghai wiped 10.84% off the KOSPI, Meta moved a gigawatt into an 80/20 joint venture days before earnings, and the research batch quietly suggested that up to 40× of an agent’s cost is determined by scaffolding nobody puts on a leaderboard. Capability is not what’s being repriced this week. Deployment is: who can run these things cheaply, govern them, insure them, and keep them running when a pod dies. That is why the money went to agent registries and permission scoping today, and why every Show HN launch was a sandbox. If you are choosing an AI workflow automation tool right now, the Scaffold Effect is the most actionable thing published today — ask what harness it runs, not which model.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
