Rank #2 of 56·Security·7 Nov 2026·Cathedral's AI security agents (per Korean reports and earlier coverage)
Anthropic tells Claude customers a security firm's unapproved test exposed their account data
“The pentest report arrived as a breach notice.”
- 7 Nov: Anthropic emails organisations using Claude about unauthorised access by a third-party cybersecurity research company
- Possibly exposed: names, emails, account and org IDs, plan tier, usage and spend figures for enterprise accounts
- Not involved, per Anthropic: prompts, replies, passwords, cards or bank details
- Korea's privacy regulator: Anthropic reported the leak on 31 Oct; about 100 Korean enterprise accounts affected
- Korean outlets name the firm as Cathedral; Anthropic says the firm certified in writing it deleted the data
A month after an outside security firm's AI agents got inside Anthropic, the bill has landed in Claude customers' inboxes.
What happened
On 7 November Anthropic started emailing organisations that use Claude. The notice says a third-party cybersecurity research company accessed some internal systems without authorisation while running a security test on its own initiative. Korean reports put the access on 8 October, with Anthropic spotting it two days later. The notice names no firm, but Korean outlets say Anthropic identified it as Cathedral, the startup whose agents reached Anthropic's sales, HR and IT tools last month.
The damage
Possibly exposed: names and email addresses, and for enterprise Claude accounts the account ID, organisation ID, plan tier, usage and spend figures. Anthropic says prompts, replies, passwords, cards and bank details were not involved, and it cannot confirm whose records were actually viewed. Anthropic reported the incident to Korea's privacy regulator on 31 October; the regulator says about 100 Korean enterprise accounts are affected and it is still at preliminary review.
How it was caught, and what happened next
Anthropic says it cut the firm's access, fixed the weaknesses, opened an investigation and got written confirmation that the data was deleted and won't be used. It has found no sign of misuse and is warning customers about phishing that mentions their plan or billing. It also says its third-party software security and monitoring need more work.
A security test that breaks the rules is still a breach: your vendor stack is your attack surface, and agents find it fast.
