South Korea Opens a Formal Privacy Probe Into Anthropic Over Leaked Claude Business Account Data

South Korea's privacy watchdog is no longer just reading Anthropic's paperwork. The Personal Information Protection Commission said on Nov 10 that it had finished a preliminary review of a data leak report filed by the Claude maker and opened a formal investigation, according to Yonhap News Agency.

The report itself came from Anthropic. On Oct 31, the company told the commission that information tied to roughly 100 Korean corporate Claude accounts had gone outside the company, Yonhap and Money Today reported. News1, in a report carried by Financial News, put a harder number on it, citing the commission: 112 domestic business accounts were named in the filing.

Read that figure carefully. Money Today stressed that it is a count of accounts, not a count of victims. It does not mean 112 companies, and it does not mean 112 people. The real reach inside those organizations is exactly the kind of thing a formal probe is supposed to pin down.

A security test that became a regulator problem

The leak traces back to an outside security firm getting into Anthropic's internal systems without permission, according to Yonhap. The wire service pointed to a Washington Post report from Oct 26, US time, that the security startup Cathedral used a vulnerability in third party business software to reach Anthropic's internal systems. We covered that intrusion, and the fight over who broke which testing rules, when Cathedral said its AI agents got inside Anthropic in under 12 hours.

What is new is what landed on Korean customers' desks. According to Newsis, also carried by Financial News, Anthropic's notice to some customers said an outside security firm ran a security check without the company's authorization. The notice said business contact details such as names and email addresses may have been exposed, along with enterprise Claude account and organization IDs, plan tiers, usage and spending figures. Newsis reported that the same notice said Anthropic had reported the incident to Korean authorities, though it did not name the agency.

So this is more than a stray contact list. By the notice's own description, the exposed fields could show which business customers pay for Claude, on what plan, and how much they use it.

Anthropic's line is that the damage stops there. Money Today reported the company's position that conversations with Claude and payment information were not exposed. It cited Anthropic technical researcher Lydia Hallie, who said on X that Cathedral reached account information such as names, emails and plan details, but not the prompts users typed, the AI's answers or payment data. Newsis dated that post to Nov 8.

The company also says it moved fast. Anthropic blocked the affected systems on the day it learned of the incident and fixed the vulnerability with the software vendor, Money Today reported, and it says it is improving security checks and monitoring for outside software. News1 reported that Anthropic also obtained written confirmation from the firm that it had deleted the data it accessed and would not use it.

None of that is a verdict, and neither is this probe. Opening an investigation is not a finding that Anthropic broke Korean law. A commission official told News1 the agency will look at the possibility of leakage by the outside security firm described in the filing and at whether any violations of the Personal Information Protection Act occurred. Money Today said the exact scope of the leak, how it happened and whether Anthropic's privacy safeguards were adequate are likely to be the central questions.

Here is the uncomfortable part for any AI lab. If the reports hold up, the weak spot was not Claude. It was the ordinary business software around it, and Korean enterprise customers are now the ones waiting to learn exactly what slipped through.