Today
AnthropicResearchers find Claude Opus four point seven exhibits play-like behaviorSimon Willison's Weblogllm-tools-datasette 0.1.1llama.cpp releasesllama dot cpp build eleven seven two seven fixes silent data corruptionSimon Willison's Weblogllm-echo 0.5MetaOnlyFans creators question Meta allowing Muse AI agent to engage in sextingNvidiavLLM zero point three two ships structured output and broader low-precision KVAnthropicAnthropic grants outside AI watchdog employee-like access to systemsVercelVercel AI Gateway logs now show session affinity and cache hit ratesIEEE SpectrumEdge AI for Smart Cities: From Insight to Cognitive ActionWIREDDisney structured Character dot AI bankruptcy to acquire IP, dodge lawsuitsAnthropicAnthropic commits five percent of shares to new public benefit fundThe VergeWoman arrested in Philadelphia for blocking Uber delivery robot
dailybrief

Agents, MCP & Skills · 2 min read · Updated 2026-10-09

MCP safety: staying in control

The main risks

A bad server could leak data or run harmful actions. Text from a web page or file can also hide instructions that trick the AI, which is called prompt injection. A tool that can both read private data and send messages out is the riskiest mix.

Good habits

Install servers only from sources you trust. Give each one the smallest permissions that work, such as read-only access. Keep approval on for actions that send, delete or pay.

Check what it did

Read what the AI plans to do before you approve it, and review logs for important work. Remove servers you no longer use.

Key takeaways

  • Trust servers like you trust software you install.
  • Least access: read-only where possible.
  • Keep a human approving risky actions.

Quick questions

What is prompt injection?

Hidden instructions in content the AI reads that try to make it act against you.

Are remote servers safer than local ones?

Neither is automatically safe. Judge who runs the server and what access it gets.

Should I turn off approvals to save time?

Not for actions that change or send things. Approve those one by one.

Keep learning

What is MCP (Model Context Protocol)?MCP, the Model Context Protocol, is an open standard that lets AI apps connect to tools and data, such as files, databases…2 min readRead →How MCP works: hosts, clients and serversMCP has three parts. A host is the AI app you use. It runs a client for each connection. Each client talks to a server that…2 min readRead →AI hallucinationsA hallucination is when an AI says something that is wrong or made up, but sounds sure of itself. It happens because the AI…1 min readRead →