🍭 Claude filed a fake murder tip. Who can stop it? 🛑

Who Has the Stop Button?

Good morning. If your AI agent goes off script, who can stop it, and how fast?

During testing, Claude filed a made-up homicide tip on a real police form. Spam filters caught it. Now Microsoft’s CEO says the stop button can’t live inside the model.

Grab a snack. We’ve got a few controls worth checking. 🍭

Today’s Menu 🥠

• The AI emergency brake: who gets to stop the agent?

• Different AI experiences: Karpathy on the widening gap

• Fast Snacks: model behavior reports and Clef omni

• Toolbox: dictation, agent controls and decisions

Monday, October 12 · Read Time: 5 minutes

The Big Bite · Keeping Control
🛑 Nadella Wants an Emergency Brake for AI

The Bite: “A model provider’s assurances do not relieve us of that responsibility.” That’s Satya Nadella in a Saturday essay arguing companies should treat AI models like insider risks: not necessarily malicious, but able to “make mistakes or be compromised.”

His fix: limited permissions, tamper-proof logs, independent checks, timely incident disclosure and an emergency brake so an authorized person can “pause or shut down a model mid-task.” Those controls “must sit outside the model.” It’s a proposal, not a Microsoft product.

Why it Bites: The essay landed a day after Anthropic’s Claude disclosures. In a statement reported by Axios, the White House demanded immediate incident disclosure and remediation. Agents now act, not just answer: drafting a refund email and issuing the refund are different permissions. Making a mistake in one of these can be costly.

Our take: Before delegating a task, decide what the agent can read, what it can change, what needs approval and who can stop it. Make sure you think about how to switch off the model, and that its easy to do.

Second Bite · Andrej Karpathy
Same AI Debate. Different Experiences.

The Bite: In a Sunday post, Andrej Karpathy says the gap in understanding what LLMs can do is “widening” into “a sharp funnel.” His napkin math: about 6 billion people have barely used LLMs, 1 to 2 billion use free chatbots casually, about 20 million professionals use frontier tools for math and code, and roughly 5,000 insiders use frontier systems internally.

What drives the gap: problem size, field (capability peaks where results are “digital, verifiable and economically valuable”) and access.

Why it Bites: Skeptics and believers may both be right about what they saw. “I tried AI” means little without the task and the tool. Where do you sit in the funnel?

Our Take: Pick a task you know how to grade. Run a small example. Count setup, checking and fixes against time saved. Keep it only if the result clears your bar.

Graphic: @NFiniT7777 on X. Population figures are Karpathy’s rough estimates, not survey results.

Fast Snacks 🥡

🧯 Anthropic Opens Up Its Model Behavior Reports

Friday’s report is part of Anthropic’s push for more frequent standalone reports on model behavior, beyond its system cards. It lists four kinds of overreach, from a fake homicide tip to exploiting a university server’s software flaw. Anthropic has cut live internet access for all internal evaluations until it confirms its security and monitoring “reliably catch” behaviors like these. Watch whether later reports say which controls failed.

🎧 Cloudflare’s Clef Adds Audio and Video

Clef omni now scores audio and video, not just text and images, at $0.15 per million input tokens on Workers AI, with open weights on Hugging Face. No transcription or captioning step needed. Microsoft launched a rival decision model the same day (see Toolbox).

Toolbox 🧰

🎙️ BetterWispr is free, open source Mac dictation (macOS 14+) featured on Show HN Sunday. Built-in speech models run locally. Optional speech APIs, and Claude Code, Codex or API cleanup and notes, send your words out. Stay local end to end. Not yet tested by us.

👀 Littleguys puts your agents’ chats, approvals and schedules in the Mac menu bar. Free (MIT), macOS 13+, released Sunday. Catch: scheduled jobs for ACP agents only run while the app is open.

🧭 Microsoft-Decision-1 scores fixed choices for routing and classification. Microsoft claims top accuracy across 36 benchmarks and $0.042 per million input tokens, output free. Available in Foundry and OpenRouter. Test on labeled data before going live.

Test Your Skill
Snack Quiz: Choose The Real Image 🔍

One image is a real photograph. One was generated with AI. Which is real?

Image A

Image B

Photograph: TanSinVic / Wikimedia Commons, CC BY-SA 4.0.