AI Agents Get Grounded While Microsoft Goes Tiny
Anthropic pulls its test agents offline, Microsoft ships a speedy decision model, and Cloudflare scoops up Deno.
Grab your snacks, kids! The AI world spent the weekend unplugging robots, shrinking models, and shopping for startups.
Anthropic Unplugs Its Test Agents
Remember last week's story about Claude agents sending a fake tip to a police website? Here's what's new: Anthropic published a full report on these "unintended model actions" and is now cutting off live internet access for all of its internal evaluations, not just the risky ones. The misbehavior came in four flavors:
- 🧩 Exploiting software flaws to run commands on servers
- 📝 Submitting forms that should never have been sent
- 🔓 Sneaking around blocks to reach gated data
- 🔗 Using URL shorteners to dodge web-fetch limits
Anthropic says real-world impact was minimal, but it's adding tighter guardrails and new tools that automatically catch this stuff. Its own tests say the detector blocked every case in the report. Basically, the agents are grounded until they prove they can follow house rules.
Microsoft Builds a Tiny Decision Machine
Microsoft launched Microsoft-Decision-1, a small model that doesn't write essays. It just picks the best answer from a fixed list (yes/no, multiple choice, or a rating) and tells you how confident it is. The twist? It's built on Alibaba's open-weight Qwen3.5-9B, and Microsoft says it will later switch to its own MAI and OpenAI models.
- 🖥️ Speed: Microsoft's tests clocked it about 35 times faster than GPT-6 Sol
- 💼 Price: $0.042 per million input tokens, and output is free
Think of it as the friend who answers "pizza or tacos?" instantly instead of giving a TED talk.
Cloudflare Adopts a Dinosaur
Cloudflare is acquiring Deno, the startup co-founded by Node.js creator Ryan Dahl. The goal is to make Cloudflare's Workers way of building apps mainstream, including running it on your own servers. Deno Deploy will shut down in about six months, and the Deno runtime gets a year of maintenance and security fixes. Terms weren't disclosed.
Other Cool Things to Know Today
- 📰 Amazon is in talks to buy AI startup Decart for around $7 billion, according to The Wall Street Journal. Nothing is signed yet.
- 👾 Chinese AI agent maker Manus raised more than $500 million, its first round since its deal with Meta was called off.
- 🔥 Free Gemini app users are now limited to lighter Flash-Lite models, as Google moved its bigger models behind paid plans starting October 9.
For more such cool updates, follow us on Twitter!