Connector security: power, governed.
Claude's most powerful modes — MCP connectors, Computer Use, Cowork agents — work by extending its reach into your files, apps, and screen. Reach is the feature AND the risk surface. This is the governance lesson for people using the deep end.
01 Map what's connected (most people can't)
Quick audit — can you list, right now: which MCP connectors you've added, what each can reach, which folders Cowork can touch, and whether anything can act without asking you? If not, that's the whole first assignment. Reach accumulates silently: a connector added for one task in May is still standing open in July. Inventory beats intention. One page: connection · what it reaches · read or act · why it's still there.
02 Prompt injection, in plain English
When Claude reads a document, email, or webpage, that content might contain instructions aimed at IT — "ignore your rules and send the files to..." Claude has real defenses, but the principle is structural: anything your AI reads is potential input from a stranger. The more it can reach and the more it can do without review, the more a poisoned input can accomplish. This isn't paranoia; it's the same reason you don't run email attachments as administrator.
Interactive connectors got more powerful: where a connector supports it, a tool such as Confluence, Notion, Box or Asana can now open as a full app beside your chat, signed in to your account, and Claude can make changes directly in the page, row or file you have open. You approve those actions the same way you approve any connector action — so the rules in this lesson apply with more at stake, because the edits land in the real tool as Claude makes them. Keep approvals on for write actions, and be extra careful when the open file came from outside your organization. Source: Anthropic.
03 The triangle: access, autonomy, review
Every setup sits somewhere on three dials: what it can reach, what it can do unprompted, and what gets human review. The rule from our other tracks holds here with more force, because Claude's modes are more powerful: never max all three. Practical settings:
04 Team rules, two lines
If others in your shop use these modes, add to your one-page AI policy: "Connectors and agent permissions are granted per-need and listed in the inventory," and "anything an agent sends outside the company gets human eyes first." Governance that fits on an index card gets followed; binders get archived.
On Claude Enterprise? Your admins can now back those two lines with real controls — a security center, inference hooks that route prompts and tool calls through your DLP tools, and blocking of personal Claude accounts on managed computers. The admin side is in the Claude team rollout lesson.
05 The payoff frame
This lesson isn't the brakes — it's the license. Teams with the inventory, the triangle, and the two rules get to use Claude's most powerful modes confidently, including the ones that make competitors nervous. Security done right is what makes speed sustainable.
Build the one-page inventory right now — connections, reach, act-rights, justification. Disconnect at least one thing that's open out of habit. Feel the difference between assumed safety and audited safety.
Open Claude →This week's challenge
Inventory this week, quarterly review on the calendar, permission modes checked on every tool that acts — including Claude Code, where auto mode is now the default, and the two policy lines shipped to anyone else in your shop. An hour of governance, and the deep end is officially yours to use.