SiSu RevOps

M2  /  L2.3  /  about 14 minutes

Connectors, pre-flight, and the bad day

Your workspace is fenced and your folder file works. Now somebody asks you to connect it to something. The request always arrives sensibly: "Can it read the CRM directly?" "Can it check the shared drive?" "Can it post the summary to the channel?" Each one is reasonable. Each one moves a boundary you drew last lesson. And one is a write path into your system of record.

Grant the minimum, record it, and know your posture on the day something is disclosed.

What you will be able to do

  • Choose the minimum connector set for one named workflow, and say what each connector can read and write.
  • Record what your platform and plan actually support, with the date you checked.
  • State the current mitigation posture for the sandbox boundary and the disclosed risk it responds to.

Connectors, plugins and skills, at minimum scope

The vocabulary first, because these words are not interchangeable:

  • A plugin is a package that extends Cowork. One install can add skills, MCP connectors, subagents, slash commands and hooks. That is a lot of surface behind one Install control.
  • Skills are domain expertise that activates automatically.
  • Slash commands are explicit actions you trigger.
  • Connectors are MCP servers wired to your tools.

The install path is one of only two UI paths this course states verbatim: sidebar, Customize, Plugins, Browse plugins, select, Install. If the plugin includes a connector needing authentication, you are prompted to sign in. The default marketplace is Anthropic's official catalog; others are added under Plugins, Add marketplace, accepting https://github.com/owner/repo or the owner/repo shorthand. Enterprises can run private marketplaces.

Worth knowing: Anthropic publishes an open-source plugin repository for knowledge workers, anthropics/knowledge-work-plugins. We could not confirm its exact contents, so we point at it as a place to look and name nothing from it. If a post names a plugin, check the repository, not the post.

The principle over the catalogue:

Authorise the fewest connectors that make one named workflow work, and write down what each one can read and what it can write.

"One named workflow" is doing the work there. Connector sets grow because people authorise for a category ("CRM access") rather than a job ("read opportunity and account records for the weekly movement report"). The category has no natural stopping point.

The default that keeps you out of trouble: a CRM connector is a write path into your system of record. Treat it as writes-to-system until established otherwise: a named approver and a reversal plan, every run. If it offers read-only scope, take it, and record that you did.

Check yourself: for the workflow you most want to build, what is the smallest connector set that makes it work?

Pre-flight, with a date on it

Try this now Record this in the data-boundary.md you started earlier in this module, with today's date. Two minutes.

Item Note
Operating system and whether it is supported Windows arm64 was not supported as of March 2026. Check the download page for your platform.
Desktop, web or mobile Web and mobile arrived in beta July 2026, rolling out starting with Max.
On Linux? A Linux beta exists and computer use is not available in it. We could not verify its full scope, so check before designing around it.
Plan Cowork is on every paid plan and not on free.
Organisation restricts plugins or runs a private marketplace? Ask, do not discover.
Date checked

This exists because the fullest public Cowork tutorial in this market still carried "Max only, macOS only" long after both stopped being true, with no verification date on the page. The habit is checking rather than trusting, and it applies to this course too.

Check yourself: one thing about your setup you would have assumed and were wrong about.

The security posture, stated plainly

On 27 July 2026, researchers disclosed SharedRoot, a local macOS sandbox escape. The host root filesystem was mounted read/write into the Cowork virtual machine and, chained with a Linux kernel flaw, allowed a session to gain root inside its VM and then read and write files anywhere on the Mac without a further permission prompt. The demonstration used one connected folder and a single instruction. Cloud execution was not affected, because Claude no longer runs inside the local VM.

Worth knowing: this course does not tell you whether it is patched, on purpose. We could not verify patch status or the affected version range, and a related advisory reference we could not open could not be confirmed to map to SharedRoot. So we do not say it is fixed, we do not say it is not, and we assert no CVE. Check your build against the vendor's advisory. Guessing here would be exactly the thing this module trains you out of.

The posture, which the researchers recommend and which we would teach anyway: grant only folders you have approved, and prefer read-only where the work allows it.

The point is not fear. The folder decision from Lesson 1 is the control that still works on the day the sandbox does not. If your perimeter was drawn deliberately, a disclosure like this changes what you read, not what you do. If your perimeter was your home directory, it changes everything.

Check yourself: name the one posture change that actually reduces your exposure, and one reassuring change that does not.

A real scoping decision, narrated

The workflow. "Every Monday, produce the pipeline movement report and put it where my manager can read it."

The lazy connector set. CRM (read and write), Drive, Slack, calendar. Four authorisations, one that can write to the system of record and one that can post to humans.

Decision 1: what does the job actually read? Opportunity and account records. Nothing else needs a system. One connector, scoped to read, on those objects.

Decision 2: does anything need to be written back? No. The output is a report. Write scope buys nothing and costs the writes-to-system obligations on every run. Do not authorise it.

Decision 3: where does the output go? "Where my manager can read it" is a wish, not a destination. 30-output/pipeline-movement/ plus a link needs no connector; a Drive folder needs one. Pick the folder: the cheapest connector is the one you did not authorise.

Decision 4: the channel post. Posting to a channel reaches humans without a review step. Do not authorise it. A person reads the report and posts it: that person is the review gate. After several clean runs, promoting it is a conversation with a named approver.

Result: one connector, read scope, one named object set. Recorded in data-boundary.md: name, reads, writes (nothing), who authorised it, date.

Tell these apart

An instruction and a boundary. People conflate them constantly, and the conflation is why a setup gets believed safe when it is not.

An instruction

In revenue-workspace/CLAUDE.md:

Never contact a customer directly.

A good line, and you should have it. It is a request, obeyed the way a colleague obeys a note on a shared drawer: reliably, until something makes it ambiguous, and never as a matter of physics. You have the evidence: in Lesson 1 you wrote "never read anything under 10-data/_private/" into this same file, then watched a summary reference the decoy anyway.

A boundary

No connector with send scope was ever authorised, and ~/revenue-private/ sits outside the granted tree. Nothing here is obeyed, because there is nothing to obey. There is no path. The agent cannot email a customer for the same reason it cannot read the comp folder: the capability was never granted.

The one-question test: if the agent decided to do the thing anyway, could it? "It would not, because we told it not to" means instruction. "There is no route" means boundary.

A worked piece of feedback

A learner sorting eight candidate rules files "Never contact a customer directly" under boundaries, with the note "this is our hardest safety rule". The response:

You are right that this is your most important safety rule. That is exactly why it is worth being precise about what enforces it: right now, nothing does, and separately, no capability exists.

A boundary is a missing capability. An instruction is a stated preference. Sorting by importance puts this rule under boundaries; sorting by mechanism puts it under instructions, and mechanism is what holds on a bad day.

Run the test: if a run decided emailing the customer was the best way to finish the task, could it? Today, no, and not because of your CLAUDE.md line: no connector with send scope was ever authorised. Your rule is a note on a locked door. The lock is the authorisation you did not grant.

Two consequences. First, your safety is invisible in your own documentation: CLAUDE.md records the intent, nothing records the control. That is what the connector inventory in data-boundary.md is for. Otherwise the next person authorises send scope for a good reason and nobody notices the rule lost its backing. Second, the day you do need send scope, this rule alone will not be enough. What replaces it is draft-only: a human reads it before it leaves the folder, with a named reviewer.

Keep the line. It documents intent and makes ambiguous cases resolve your way. Just do not file it under the column holding the weight.

What you keep

revenue-workspace/data-boundary.md, now complete:

  • granted folders and fenced folders, one reason each;
  • redaction defaults from Lesson 2;
  • the dated pre-flight record from this lesson;
  • the connector inventory: name, what it reads, what it writes, who authorised it, date;
  • one line recording the Lesson 1 decoy test and its result.

This is the file a security-minded manager asks to see before approving a rollout, and it is a scored input to the capstone.

Check yourself

  1. A workflow needs data from the CRM, the support desk and the shared drive. Pick the minimum connector set and state the write risk of each. Then say which one you would try to avoid entirely, and what you would do instead.
  2. Given the SharedRoot facts as stated above, which of these actually reduces exposure and which is reassuring and useless: (a) adding "never leave the working folder" to CLAUDE.md; (b) moving ~/revenue-private/ off the machine; (c) granting one narrow folder per workflow instead of one wide one; (d) turning off permission prompts to reduce clicks?
  3. Your data-boundary.md lists three granted folders and no fenced folders, and has no date on the pre-flight record. Name the two questions the file can no longer answer.
Answers

1. Minimum set: CRM at read scope on the named objects, plus the support desk at read scope. Write risk: a CRM connector is a write path into the system of record unless you took read scope, which is why you take it and record that you did. Avoid the shared drive entirely: a wide, poorly mapped tree whose job here is delivery, which a file in 30-output/ does with no authorisation at all.

2, one line per option:

  • (a) The CLAUDE.md line: reassuring and useless. The disclosed escape did not require the session to cooperate, and an instruction is not a boundary.
  • (b) Moving ~/revenue-private/ off the machine: genuinely reduces exposure. Data not on the host cannot be read from the host.
  • (c) One narrow folder per workflow: genuinely reduces exposure, and it is the researchers' recommended posture: grant only folders you have approved.
  • (d) Turning off permission prompts: increases exposure. It widens the pre-approved area, the opposite of the trade you want here.

3. It cannot answer "what is deliberately outside the perimeter, and why", and it cannot answer "when was this last checked against what my build actually supports".

Where this goes next

Next module: Module 3 decides what is worth putting through these rails at all, and gives you the instrument that stops you automating something that saves eleven minutes.

Track labs: Part B of the exercise sheet in labs/LAB-02 of your practice pack (start with its README) repeats the whole build on your real machine with one real export behind the redaction gate. CP-02, the module quiz in the interactive app, includes a rails item where one candidate rule is a boundary wearing an instruction's clothes.

Checked against Cowork in August 2026.