Anthropic's $25B, MCP SDK RCE, and Kimi's 12-Hour Horizon

Top-Line Summary

Amazon committed another $25B to Anthropic ($5B now, $20B later) in exchange for Anthropic pledging $100B on AWS Trainium and 5 gigawatts of power - and the same week, a security firm disclosed RCE across the official Anthropic MCP SDKs with ten CVEs and Anthropic’s response was that the behavior is “as expected.” Opus 4.7 landed, the hosts are split on whether it’s smarter or just more verbose, and Kimi K2.6 quietly pushed long-horizon agent work to 12 hours and 300 parallel sub-agents.

Show Video

The “Pre-Show” Context

Another Wednesday night, another sauna of an office — it’s the third week of April in Ontario and Brett and Travers are finally out in their backyards without four layers on, with snow forecast again next week because there is no such thing as spring here. Brett’s Mac reset his mic input to 50% again mid-cold-open, so Descript is going to earn its money on this one. They also detoured into their favorite genre (“so bad it’s good” movies): Chain Reaction with Keanu Reeves inventing a new energy source, Johnny Mnemonic, Killer Sofa, and the one where the tire rolls through the desert killing people.

The Engineering Rundown

• Amazon to invest up to another $25B in Anthropic (03:00)

  Another $25B on top of the existing $8B — $5B up front, $20B contingent on “commercial decisions,” which Brett reads like a professional sports trade with future considerations. In exchange, Anthropic commits $100B on Trainium and 5 gigawatts of power, which is awkward because the same week Jassy’s shareholder letter admitted AWS is leaving revenue on the table for lack of electricity. Travers’ tangent about deploying Workspaces because corporate laptops just jumped $1,500 a unit is the more useful engineering datapoint here — hardware costs are inflating because the hyperscalers are consuming everything. Brett’s prediction: we’re walking back to the mainframe era, with dumb terminals and rented compute, and cheap 64GB of RAM is not coming back. Anthropic has also openly admitted its infrastructure is under “inevitable strain” with reliability issues, which tracks with the chatter in power-user Discord channels.

• AWS Lambda can now mount S3 buckets as file systems (13:35)

  Multiple Lambdas can now mount the same S3 data as a file system and do standard file operations without downloading anything first. It’s built on the same primitives as EFS, so the mental model is the EFS pattern — one file system, thousands of consumers — applied to S3. The obvious fit is fan-out batch processing where you’ve been stitching S3 and EFS mount targets together by hand; you can stop doing that now.

• Aurora Serverless: 30% Better Performance, Still Scales to Zero (14:23)

  A drive-by AWS item Travers spotted on the news feed: Aurora Serverless now claims 30% better performance with smarter scaling, still scales to zero. Brett’s speculation: 30% is the number you hear when an AWS service moves to Graviton underneath, so the architecture may have quietly shifted. Either way, free performance on a service you’re probably already defaulting to is a nice pickup.

• Travers’ NLP Topic Map on the Podcast Back-Catalog (16:20)

  Travers kept going on his local-Whisper + ROCm transcript pipeline from last week — now building a topic map of everything the show has covered so he can surface follow-up topics and callbacks. He also stepped off vibe-coding for a few days and is grinding CodeWars problems (Huffman encoding at the moment) to keep the hands-on-keyboard developer muscles from atrophying. Both hosts admit the break only lasted a weekend.

• Brett’s Kanban and the “What Now” Agent (19:56)

  Brett is obsessed with his Kanban board again — he’s wrapped it in an MCP so agents can read and write cards, bound cards to AWS accounts via an account-finder tool, and added deep links that drop him straight into the right account for the work. Over the weekend his brain surfaced the real question: it’s still just a Kanban board. If he’s the bottleneck moving cards between columns, why is he the bottleneck? The direction now is a “What Now” agent that periodically scans the board, classifies cards as easy-enough-to-do-itself (documentation updates are the canonical example) or human-required, and handles the easy ones. Implementation: ECS Fargate container, not Bedrock AgentCore, after the model tried to gaslight him about AgentCore capabilities and he made it go read the docs.

• Introducing Claude Opus 4.7 (29:07)

  Pitched for “complex, long-running tasks.” Brett hasn’t noticed much difference from 4.6, but a lot of Discord chatter says 4.7 is “dumber than before” — Travers has been hearing the same. Brett flipped to Sonnet 4.6 for a few days mid-last-week because of the token-churn scare stories, noticed the model did struggle briefly, then switched back to 4.7 and things settled down. The one concrete change: new “effort control” with a default of xhigh (between high and max), which lines up with the ChatGPT verbiage. Output is noticeably more verbose — some of it useful, some of it a bit LLM-cringe (the model is calling conversations “convoys” now and keeps asking “want me to sketch this out?”). Travers’ working theory: there may be a router model sitting in front of Opus deciding which variant answers, and the GPU constraints Anthropic keeps flagging could explain the inconsistency people are reporting. If 4.7 feels off this week, you’re probably not imagining it.

• The Architectural Flaw at the Core of Anthropic’s MCP (37:20)

  This one is bad. Remote code execution on any system running a vulnerable implementation of the official Anthropic MCP SDKs — the bug is in the SDKs themselves, so if you’re building on top of them you are exposed. Ten CVEs so far across light-llm, GPT Researcher, LangChain Chat, and others. Potentially 200,000 vulnerable instances in the wild. The punchline: the security firm notified Anthropic, and Anthropic “declined to modify” saying the behavior is “as expected.” The remediation list is long but the one Brett keeps coming back to is the old AMI rule — use official, trusted directories only. Community MCPs are now firmly in the “USB key on the sidewalk” category, and Simon Willison is already calling that a disaster waiting to happen. Brett’s own mitigation: he rebuilds every Docker MCP Toolkit image from scratch rather than pulling whatever’s in the repo.

• Codex for (almost) everything (45:10)

  Codex can now drive every app on your computer (including clicking), ships an in-app browser useful for frontend and game dev, has a 90+ plugin library, and leans hard on new automations. Brett is still treating Codex like the black sheep of the toolbox — it’s in the dock, he gets to it later. Travers has flipped: Codex is his go-to after OpenCode right now, and Claude Code has slipped because Codex just feels more consistent for the work he’s doing. The meta-takeaway the hosts keep landing on: don’t get sticky on one tool. The field is moving fast enough that you need to be fluent across Codex, Claude Code, OpenCode, Gemini, and whatever ships next month, or the day your favorite tool changes you’ll be stuck starting over. That’s part of why a second-brain/OpenBrain layer outside any single tool matters — context portability is the hedge.

• AI chip startup Cerebras files for IPO (52:07)

  Cerebras (of “big chips” fame) filed for IPO — no date, no valuation yet, and this is their second attempt. They’ve got deals with both AWS (chips going into AWS data centers) and OpenAI (training). Whether this is peak demand or mid-demand depends entirely on whether the power gets built, which circles right back to the $25B Anthropic deal.

• Google’s Eighth-Generation TPUs for the Agentic Era (52:56)

  Two chips announced at Google Next: 8t for training, 8i for reasoning. Ties directly to the Cerebras story — the “one GPU for everything” era is over, and everyone is splitting silicon by workload type. Google’s been building TPUs a long time, and a huge chunk of their compute infrastructure sits on that stack, which Travers thinks is a real advantage as long as Google can ship a developer experience that isn’t Gemini-tier painful. Brett is using Google Antigravity as a backlight right now, which is probably not in the telemetry slide.

• Cloudflare’s Internal AI Engineering Stack (56:44)

  Every LLM request at Cloudflare goes through a single AI Gateway Worker — one place for provider keys, cost tracking, data retention. Steal this pattern if you’re rolling your own. The operational number that matters: 7 billion tokens a day running on Kimi K2.5 via Workers AI, 77% cheaper than frontier-model equivalent. That’s the real data point for the “do I actually need Claude for this?” conversation — most tasks don’t need a frontier model, and if you’re on a subsidized subscription you don’t feel the cost signal until the day the subsidy changes. The other thing worth the deep-dive read: every MR gets a multi-agent code review (security, quality, performance, compliance) that cites their internal Engineering Codex. That’s what AI-enforced coding standards actually look like in production.

• Moving Past Bots vs. Humans (1:01:40)

  Cloudflare’s argument: the bot-vs-human split is getting outdated. What matters is intent and behaviour, not the category, and the path forward is privacy-preserving credentials (Privacy Pass, ARC, ACT) that prove you haven’t blown your rate limit without revealing who you are. The stakes are the open web itself — without something like this, more sites end up behind logins or paywalls, which is the direction we’re already headed. Ties into Cloudflare’s CDN spec for raw-HTML delivery that’s more token-efficient for agents.

• Podcast Metrics, IAB Standards, and Spoofed User-Agents (1:03:34)

  Brett killed his podcast metrics SaaS and rebuilt the whole thing on S3 + CloudFront for pennies a month. Once a week he asks Claude to analyze the CloudFront logs and filter real listens from bots using the IAB standard — which measures listen-time by bytes downloaded, not requests. The surprise find: a chunk of apparent Apple Podcasts traffic had user-agents that didn’t match Apple’s documented string format (e.g. a version number that isn’t in the published standard), and Claude flagged those as spoofed bots. This is a useful exercise for any AI-assisted log analysis project — “use the official standard, here are the valid values, filter what doesn’t match” is a repeatable pattern.

• Kimi K2.6 Tech Blog (1:10:00)

  Long-horizon is the headline: 12+ hours on a single task, 4,000 tool calls, 14 continuous executions — they had it download Qwen 3 0.8B locally and optimize the inference in Zig, going from 15 to 193 tokens per second (20% faster than LM Studio). Last week Brett and Travers called 8-hour horizons the current ceiling; now it’s 12. Also scales horizontally to 300 sub-agents. If you’re building anything agentic, long-horizon + massive parallel sub-agent fan-out is the shape that keeps showing up, and Kimi is worth putting on the bench alongside your frontier default.

• Multi-Agent Code Review, Stolen from Cloudflare (1:13:45)

  Brett’s takeaway action item from the Cloudflare post: wire up a second model (Gemini or Codex) to review every merge request that lands in his GitLab before it gets to him. His current flow is already agent-driven — the agent decides whether a change is safe to self-merge (docs, small updates) or needs to be gated to Brett (infrastructure changes like the metrics DynamoDB table). Adding a review agent in between is the factory-line extension of that: production cell turns out PRs, review cell fires multiple agents with different toolsets and contexts before it ever hits a human. Expect a working prototype by next week.

Off-the-Clock Recommendations

• Chain Reaction (1996) and Johnny Mnemonic (1995) — both Keanu. The first has him inventing fusion-adjacent energy; the second has him overloading his brain’s storage. Classic “so bad it’s good” genre.
• Killer Sofa (2019) and Rubber (2010) — the tire-rolling-through-the-desert-killing-people one. Yes, really.
• Red Rising by Pierce Brown — Brett is 70% through the first book and binge-reading to 1am. He calls it Silo + Matrix + Dune + Firefly vibes without giving anything away: humans have conquered the galaxy, terraform planets, fly interstellar ships, and still fight with swords.
• Go outside. The Simon Willison recommendation. Third week of April in Ontario, the weather is finally cooperating.