South Korea's privacy regulator reviews Anthropic breach that exposed about 100 corporate accounts
Key Takeaways
- •South Korea's PIPC received Anthropic's breach report filed on October 31 and is conducting an early-stage review of an incident that exposed about 100 business accounts belonging to Korean companies.
- •The exposed data included account and organization ID numbers, business contact details such as names and email addresses, and business subscription plans, while user prompts, AI outputs, payment data, model weights, source code, and Claude conversations were not accessed.
- •Anthropic shut down the affected systems the same day it confirmed the problem and worked with the third-party software vendor to fix the exploited flaw, while acknowledging that more work remains on monitoring third-party software.
- •The breach is linked to Cathedral, a cybersecurity startup founded by four former DOGE staffers and valued at $1.4 billion after raising $160 million from Andreessen Horowitz and Sequoia Capital, which claims its AI agents reached internal Anthropic data in under 12 hours, a timeline Anthropic's cybersecurity lead disputes.
- •Anthropic says Cathedral broke its bug bounty rules by not fully explaining the test's scope and overstating its national security significance, and no legal complaint has been filed while Cathedral has entered an agreement to test OpenAI's defenses.
South Korea's privacy regulator is reviewing a data breach at Anthropic that exposed roughly 100 business accounts belonging to Korean companies.
The Personal Information Protection Commission (PIPC) said on Monday that it received a breach report from Anthropic, which the company filed on October 31. The regulator's disclosure was reported by Yonhap News Agency and Seoul Economic Daily. The PIPC serves as South Korea's top personal data-protection authority, and the review places an overseas AI provider's handling of Korean corporate customer data under official scrutiny.
What did Anthropic report to the PIPC?
The PIPC confirmed that Anthropic filed a report regarding the data breach on October 31, stating that about 100 business accounts — all of them belonging to Korean companies — were exposed.
The review is at an early stage, and the commission has not yet opened a formal investigation. It noted that reviews involving foreign firms can take longer, because regulators may request documents before an investigation formally begins. Once a formal investigation starts, the PIPC said it plans to confirm the scale of the damage. The case also illustrates how Korea's data-protection oversight reaches foreign AI providers whose Korean corporate customers are affected by a breach.
Anthropic told some of its customers that the exposed data included account and organization ID numbers, along with business contact details such as names and email addresses.
Lydia Harley, a technical researcher at the company, disclosed on X that the intruder also viewed business subscription plans, though user prompts, AI outputs and payment data were not accessed. Anthropic's model weights, source code and Claude conversations were likewise not reached.
The company said it shut down the affected systems the same day it confirmed the problem and worked with the third-party software vendor to fix the exploited flaw. Anthropic has also acknowledged that more work remains to be done on monitoring third-party software. The incident shows how third-party software can become an exposure point for AI providers even when their core systems stay untouched.
Who is Cathedral, and how did the breach happen?
Cathedral is a young cybersecurity startup founded by four former staffers of the disbanded Department of Government Efficiency (DOGE): Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. The company is valued at $1.4 billion after raising $160 million from Andreessen Horowitz and Sequoia Capital, and Marc Andreessen and Shaun Maguire both sit on its board.
The startup reportedly aims to build up U.S. military cyber power against rivals such as China, and it holds contracts with three federal agencies.
Kliger, the CEO, has the closest history with Anthropic. He previously served as chief data officer at the Pentagon and, by several accounts, took part in the department's clash with Anthropic over how the military could use its AI models.
According to Cathedral's account of events — logged in MIT's AI Incident Database — its AI agents found paths to meeting transcripts, salary and job-offer details, and internal Slack messages in under 12 hours.
Logan Graham, Anthropic's cybersecurity lead, refuted that claim, posting on X that the effort actually took several days. He wrote that Cathedral's method was "unusual and out of scope" because it combined flaws from several third-party software products.
Cathedral's test of Anthropic's defenses was supposed to fall within the company's bug bounty program, but Anthropic says the startup broke the program's rules by not fully explaining how far the test went, and by making its findings sound more significant to national security than they were.
Anthropic's own monitoring systems reportedly caught part of the activity before Cathedral filed its report.
So far, Anthropic has not filed a legal complaint, and Cathedral has entered into an agreement with OpenAI to test its own defenses. Whether the PIPC's early-stage review escalates into a formal investigation remains the immediate follow-up point, alongside any decision by Anthropic to pursue legal action.
