The weekend's AI coverage was mostly about Jensen Huang again: a podcast, a long recap and a Saturday Night Live sketch. I've written about Nvidia's leadership crisis four days in a row, so I'm stopping. Sunday did produce a few things that matter more if you build software for a living, and the most useful was a single number buried in a guest column.
78.3%.
Palo Alto Networks' Unit 42 measured that attack success rate against an agent connected to five MCP servers. Today's post is ordered by what you can actually do with each story: fix it this week, watch it, file it away, or ignore it.
Fix this week: your MCP servers trust each other more than you do#
The number comes from a VentureBeat guest piece by Nilay Dhamsania, a director at the consultancy Intuz. It's a column, not original research, but the sources it compiles are solid: Unit 42's measurement, the OWASP MCP Top 10, NSA guidance from May 2026 and Microsoft's own MCP governance notes. The core idea is simple and, I think, still underrated.
Call it cascading trust. Once an agent treats one server's output as trustworthy context, a compromised server can inject instructions that carry into every later tool call. That includes calls to servers that were never compromised. Your well-audited GitHub integration becomes the weapon because a sketchy weather server told the agent what to do with it. The more servers you attach, the bigger the blast radius, which is exactly what the five-server figure shows.
The column's second point is about protocol history, and I suspect most teams haven't absorbed it. MCP's auth model has changed in nearly every spec revision:
2024-11-05: launch, with no authorization framework at all
2025-03-26: an OAuth 2.1-based framework arrives
2025-06-18: servers become OAuth Resource Servers, and Resource Indicators (RFC 8707) are required
2025-11-25: Client ID Metadata Documents (CIMD) are added alongside Dynamic Client Registration
2026-07-28: DCR is deprecated, and clients must validate the iss parameter (RFC 9207)
If you wrote an MCP server in its first year, it probably relies on assumptions the spec has since dropped. Per the NSA mapping, only two of OWASP's ten risk categories sit in the text the agent reads. The other eight, including credentials, scope and identity, live in infrastructure that MCP doesn't constrain.
"It never standardized how you prove who's calling or what they're allowed to touch." — Nilay Dhamsania, in VentureBeat
My to-do list from this is short. Pull credentials at call time instead of caching them at startup, because you can't meaningfully revoke a token that sits in process memory. Scope tokens down hard; Microsoft's guidance warns against "full mailbox access where read-only would do." Start validating iss now. And once you run more than a handful of servers, put an identity-aware gateway in front of them that pins tool definitions and alerts on drift, so a quietly changed dependency gets noticed. Per-server enforcement only works if every server author gets it right, and they won't.
Watch: Apple wants 200,000 people to test health answers before it ships them#
Apple launched the Apple Health AI Study App, a chatbot-style app built with OpenEvidence that answers everyday medical questions using clinical sources such as NEJM, JAMA and Nature. With permission, it can also read HealthKit data like resting heart rate, sleep and lab results.
Up to 200,000 participants, for more than 12 months
US only, age 21+, iPhone 15 or later, iOS 26, Apple Intelligence enabled
OpenEvidence deletes query and response data after 24 hours; Apple says none of it trains models
What interests me is the order of operations. OpenAI said earlier this year that users submit more than 300 million health queries a week, and that product was shipped first and studied afterward. Apple is running a year-long study with monthly surveys and answer ratings before saying whether any of this will reach Siri or the Health app. You can call that slow. For a domain where a confident wrong answer hurts someone, I call it the only defensible way to build an eval set. If you work on regulated AI products, Apple's design is worth copying: ground answers in a fixed set of sources, collect ratings from real users, and keep data retention short.
File away: a $24M bet that the sensor should do the thinking#
New Delhi-based Bharat Semi raised a $24M seed round led by Info Edge Group, South Park Commons and Indus Bridge Ventures. The company builds programmable image sensors that capture, process and run AI on a single chip. Co-founder Vrinda Kapoor claims that cuts system costs by at least 70% compared with a separate sensor, ISP and SoC. It already has design wins with General Atomics and Kaynes Semicon, and its CTO used to hold the same job at OmniVision.
It's also pre-revenue after nearly five years, and commercial production is targeted for 2027. I'm not drawing conclusions yet. But with GPU supply suddenly looking like a governance risk, more inference moving into cheap silicon at the edge is a trend worth tracking.
Ignore, mostly: the Huang post-mortems#
TechCrunch's Equity crew and The Verge's Stepback both recap the scandal well. The only figure I'd keep is Nvidia's exposure: Rebecca Bellan cites $99 billion in investments, $25 billion in commitments and $36 billion in cloud service commitments, spanning CoreWeave, Nebius, OpenAI and Anthropic. Anthony Ha called Colette Kress's appointment "a caretaker appointment." A CFO-led Nvidia means more scrutiny of that portfolio, and that will matter more than the petition.
Where the risk actually sits#
What links these stories is where trust goes. Apple is putting its trust in clinical sources and a year of measurement. MCP deployments are trusting whatever the last tool returned. My prediction: within six months, an MCP gateway with pinned tool definitions will be as standard as an API gateway, and the first big agent breach post-mortem will describe a cascade from a server nobody thought was important. I'm auditing my own MCP setup's token scopes this week. It's the cheapest fix in this post.