Review the third-party business apps every client connects.
An authorized test of Anthropic's systems shows where AI agents look first: the sales, HR, meeting and chat tools around the core.

Published October 27, 2026 · ~5 min read
What happened
On Oct 26, 2026, The Washington Post reported in an exclusive that Cathedral, a cybersecurity start-up a few months old, used AI agents to find routes into Anthropic's internal business systems. Cyber Security News published detailed coverage the same day.
This was an authorized test run under Anthropic's program for outside researchers, its bug bounty. It was not a criminal attack. Cathedral says its agents found the routes in less than 12 hours. That figure comes from Cathedral's draft report, which the Post reviewed.
Anthropic's monitoring flagged some of Cathedral's activity before Cathedral reported it. Anthropic shut off access the same day. It disabled the affected business software and turned it back on only after the vendors confirmed fixes. Anthropic says it found no evidence that anyone outside Cathedral used the weaknesses.
What the agents used, and what they did not
The agents did not break Anthropic's models, and they did not exploit Anthropic's own code. They used weaknesses in third-party software that Anthropic uses for administrative and sales work. The vendors and the specific flaws have not been named.
The reported access included meeting transcripts, salary and job-offer details, and internal Slack messages. Anthropic confirmed access to a limited group of sales, human resources and IT systems. Those systems held business contact information for some customers and partners, plus basic user account records.
Anthropic says these were not accessed: customer prompts and conversations, Claude responses, source code and model weights.
So the core systems held. The exposure was in the layer around them, meaning the SaaS tools a company signs up for to run sales, HR, IT and meetings, and the connections between those tools.
The claims, and whose they are
AI agents found routes into internal business systems in less than 12 hours.
Source: Cathedral draft report, reviewed by the Post. Status: Cathedral claim.
Access reached meeting transcripts, salary and job-offer details, and internal Slack messages.
Source: The Washington Post. Status: Reported by The Washington Post.
Access covered a limited group of sales, human resources and IT systems.
Source: Anthropic. Status: Anthropic statement.
No customer conversations, Claude responses, source code or model weights were accessed.
Source: Anthropic. Status: Anthropic statement.
Monitoring flagged activity before Cathedral reported it, and access was shut off the same day.
Source: Anthropic. Status: Anthropic statement.
No evidence anyone outside Cathedral used the weaknesses.
Source: Anthropic. Status: Anthropic statement.
Cathedral did not fully follow program rules because it did not disclose all its activity.
Source: Anthropic. Status: Anthropic statement.
Cathedral kept data access to the minimum needed.
Source: Cathedral. Status: Cathedral claim.
Why this matters for an MSP
Every client runs on a stack like this one. They have a CRM, an HR and payroll platform, a meeting-notes tool that records and transcribes calls, and a chat workspace. These tools connect to each other through OAuth grants, API tokens and integrations. Someone approved each of those connections once, and often nobody has looked at them since.
That layer tends to be the soft one, for three reasons. It sits outside the systems an MSP usually patches and monitors most closely. It holds sensitive records: job offers, salaries, customer contacts, call transcripts. And its access rules are set by many different vendors, each with its own admin console and defaults.
The Cathedral test points to a change in speed. AI agents shorten the time it takes to find a flaw in one app and chain it to the next one. Set aside the exact number of hours, which is Cathedral's own claim. Many practices review this layer once a year, or only when a client is onboarded. That schedule assumes anyone probing it moves slowly, and that assumption is weaker now.
The response is the second lesson. Anthropic's monitoring saw the activity early, Anthropic cut access the same day, and it kept the affected apps off until the vendors confirmed fixes. A client can only respond that way if someone has already mapped its apps, knows what normal API traffic looks like, and has agreed ahead of time who is allowed to switch an app off.
For your practice, this comes down to scope. Many MSP agreements cover endpoints, identity and the network, and leave the business apps a client signs up for unclear. If the third-party business-app layer is not in your current review, decide with each client whether it should be, and who on their side owns each app. The client owns these systems. Your job is to make sure someone is watching them and that the rules are written down.
A connected-app review to run with each client
Inventory.
List every SaaS app the client uses and every integration between them, including tools that one team adopted on its own. Record who owns each app on the client side and what data it holds.
Grants and roles.
Pull the OAuth and app grants from the client's identity provider and main platforms. Remove any grant nobody can explain. Cut admin and integration roles down to the people and services that actually need them.
Notes and exports.
Review the meeting-notes, transcription and chat tools. Check where recordings and transcripts are stored, who can read them, and which export or API permissions are turned on.
Alerting.
Set alerts for unusual API pulls and bulk exports from HR, sales, notes and chat systems. Send those alerts to a person who will look at them the same day.
Kill switch and fix confirmation.
For each app, agree how to disable it or revoke its tokens quickly, and who approves that call. Write down the rule Anthropic followed: the app stays off until the vendor confirms the fix.
Read with care
No product list.
The vendor names and the specific flaws were withheld, so there is no product list to act on. Review the whole category, not one named tool.
Cathedral's number.
"Less than 12 hours" is Cathedral's claim, taken from its own draft report.
Disputed parts.
Anthropic and Cathedral disagree on parts of the test, including whether Cathedral fully followed the program rules.
Authorized test.
This was an authorized test under Anthropic's researcher program, not a criminal attack. Anthropic found no evidence of use by anyone outside Cathedral.
This guide is Consultiply's independent reading of public reporting on an authorized security test. It is not a recommendation to buy any product.
Sources
The Washington Post, exclusive report by Elizabeth Dwoskin and Gerrit De Vynck, Oct 26, 2026.
Cyber Security News, detailed coverage by Guru Baran, Oct 26, 2026.
Reviewed, not run by us.
Consultiply's independent reading of public reporting on an authorized security test. Not a recommendation to buy any product. The Washington Post: researchers at cybersecurity start-up Cathedral got inside Anthropic (Oct 26, 2026) →