AiStaffo

AI agents reportedly reached Anthropic business systems in 12 hours

Stock photo: laptop keyboard office desk
Photo: Samer Daboul / Pexels

Cathedral, a months-old cybersecurity startup, says its AI agents found routes into Anthropic's sales, HR and IT systems in under 12 hours. The Washington Post reported the claim on 26 October 2026, citing a draft Cathedral report. Anthropic confirms that the researchers reached a limited set of those systems, which held business contact details and basic account records. It says customer chats, Claude responses, source code and model weights were not reached. Both sides dispute how far the test went and how serious it was. The entry point was weaknesses in third-party administrative and sales software, not Anthropic's AI models. For businesses that automate work, the lesson is that connected business apps are part of the security perimeter.

In short

  • Cathedral says its AI agents reached Anthropic's sales, HR and IT systems in under 12 hours, according to The Washington Post.
  • The entry point was third-party business software, not Claude or Anthropic's core code.
  • Anthropic confirms limited access to business contact and account records, and says model weights and source code were not reached.
  • Anthropic and Cathedral disagree on how far the test went and how serious it was.
  • Businesses should inventory connected apps and cut unused or excessive permissions now.

What happened

According to The Washington Post, a cybersecurity startup called Cathedral used AI agents to find paths into Anthropic's internal business systems in less than 12 hours. The Post's account relies on a draft Cathedral report it reviewed. Cathedral was founded by four former staffers of the US Department of Government Efficiency, according to Decrypt's summary of the Post report. The work was reportedly carried out under Anthropic's vulnerability disclosure programme for outside researchers, which Anadolu Agency describes as a programme that lets independent researchers check the company's systems.

The agents did not break into Claude. According to the Post's reporting as summarised by Decrypt, the way in was third-party administrative and sales software of the kind many companies use to track customers and manage staff. The draft report says the routes exposed meeting transcripts, salary and job-offer details, and Slack messages about Anthropic's advanced AI models.

Anthropic confirmed to the Post that the researchers reached a limited set of sales, HR and IT systems. Those systems held business contact details for some customers and partners, along with basic user account records. Anthropic says customer chats, Claude responses, source code and model weights were not reached. It also says its monitoring flagged some of Cathedral's activity before the startup reported it, that it shut off access the same day, and that it closed the access paths and informed federal officials.

Where the two sides disagree

Anthropic says Cathedral overstated the national security implications and did not disclose all of its testing, according to Decrypt. Cathedral says it kept its data access to the absolute minimum. Gizmodo reports that Anthropic said the researchers had earlier spent several days trying without success to breach its software directly. Anthropic's published Responsible Disclosure Policy states that it does not cover systems owned or run by third-party service providers, which is the category at the centre of this case. Anthropic also says it found no evidence that anyone other than Cathedral exploited the flaws. That does not show whether others could have.

Why it matters for businesses that automate work

Most automation projects connect a set of ordinary business apps: a CRM, a payroll or HR tool, a shared chat workspace, and a set of accounts that let software read and write data. Each connection is a door. In this case the agents did not need a flaw in the core product. They needed a flaw in one connected tool and a route from there into other systems. Automated agents can test many routes faster than a person, which is why the speed of 12 hours stands out.

What changes in practice

A business of any size can apply the same checks without waiting for a headline. The steps below are general good practice, not a fix for this specific flaw, which has not been named publicly.

  1. List every software tool that connects to your email, chat, file storage, HR or sales records.
  2. Remove access for former staff, unused tools and old test accounts.
  3. Limit each connected app to the data it needs. A reporting tool rarely needs write access.
  4. Review the permissions that connected apps hold, including the tokens that let one app act on another's behalf.
  5. Ask each software vendor how it handles security updates and whether it has published a way to report flaws.
  6. Set alerts for unusual logins or bulk data exports from any connected account.

Automation can run these checks on a schedule and produce a list of changes for an owner to approve. It cannot patch a vendor's product. The business can only restrict access, remove tools and press the vendor for a fix.

What to watch next

Three things will show how serious the findings were. The first is whether Cathedral publishes its final report, since the Post reviewed only a draft. The second is whether the vendors behind the affected software confirm fixes. Anthropic says it kept affected business software offline until vendors confirmed fixes, according to Superpower Daily's summary. The third is whether Anthropic and Cathedral resolve their dispute over scope and disclosure, which the public record does not yet settle.

How AiStaffo would automate this

AiStaffo builds AI workers for routine business administration, so the part of this case it can address is the connected-tool inventory and the reporting around it, not the security testing itself. A worker could list every app linked to email, chat, HR and sales records, then check each one for accounts that have not logged in recently, permissions wider than the task needs, and access held by former staff. It would produce a monthly report and flag each item for review. The owner still decides what to remove, and still asks vendors about their fixes. Book a free automation audit.

Questions people ask

Did Cathedral hack Anthropic's Claude AI model?
According to The Washington Post, the agents did not break Claude or exploit Anthropic's own code. The reported entry point was third-party administrative and sales software. Anthropic says customer chats, Claude responses, source code and model weights were not reached.
What data did the researchers reach?
Anthropic confirmed access to a limited set of sales, HR and IT systems, including business contact details for some customers and partners and basic user account records. Cathedral's draft report describes meeting transcripts, salary and job-offer details, and Slack messages. Anthropic and Cathedral disagree on how far the access went.
Was the test authorised?
The test was reportedly run under Anthropic's vulnerability disclosure programme, which lets outside researchers look for flaws. Anthropic disputes parts of how the test was run, including how much activity was disclosed. Anthropic's published policy does not cover systems run by third-party service providers.
Which software was affected?
The software has not been named publicly. The reported weakness was in third-party administrative and sales software that many other companies also use. Anthropic says it disabled the affected software until vendors confirmed fixes.
Should a business stop using connected business apps after this?
No single report is a reason to stop using these tools. A business can inventory its connected apps, remove unused access and limit permissions to what each tool needs. It can also ask each vendor how it handles security fixes.
Can automation prevent this kind of access?
Automation can check connections and permissions on a schedule and flag risky items. It cannot fix a flaw inside a vendor's product. A person still decides what to remove and follows up with the vendor.

Book a free automation audit

Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.

We use these details only to reply to this request and prepare your audit, and keep them for up to two years. We send no marketing unless you tick the box above. Privacy policy

ai securityanthropicai agentsthird party softwareaccess controls