Bug-Bounty Researchers Reached Anthropic Business Systems Through Third-Party Software

Summary
Anthropic confirmed Cathedral researchers accessed a limited set of sales, HR and IT systems during a bug-bounty test, reportedly via weaknesses in unnamed third-party software. Anthropic said Claude conversations, source code and model weights were not accessed.
Key points
- The Washington Post reported that Cathedral used AI agents to identify access paths in less than 12 hours, after earlier unsuccessful attempts to compromise Anthropic’s own software.
- Cathedral reportedly exploited weaknesses in unnamed third-party business applications, rather than Anthropic’s AI models or proprietary code.
- Anthropic confirmed access to a limited set of systems supporting sales, HR and IT; those systems held some customer and partner contact information and user account records.
- Cathedral reported reaching internal meeting transcripts, employee salary and job-offer details, and Slack discussions involving Anthropic’s advanced models; Anthropic did not confirm those specific details in the article.
- The activity took place through Anthropic’s bug-bounty program. Anthropic said its monitoring detected some activity, then it terminated access and disabled the affected software pending vendor fixes.
- Anthropic disputed aspects of Cathedral’s investigation, including whether all activity was disclosed under the program’s rules, and said it found no evidence that others had exploited the weaknesses.
- The article recommends inventorying third-party software, assessing its components and access, and improving exposure assessment and remediation; it also promotes ReversingLabs’ software-analysis product.
Article Details
- Victim Organization
- Anthropic
- Incident Type
- Researchers accessed internal systems through weaknesses in unnamed third-party business software during activity conducted under Anthropic's bug bounty program.
- Disclosure Date
- 2026-10-26
- Data Types Exposed
- Business contact information for some customers and partners, and certain user account records (confirmed by Anthropic)
- Internal meeting transcripts, employee salary and job-offer details, and Slack discussions involving advanced AI models (reported by Cathedral)
- Affected Records
- Not disclosed
- Affected Data Size
- Not disclosed
- Operational Impact
- Anthropic detected some activity, terminated access, and disabled the affected business software pending vendor fixes. Anthropic said customer conversations with Claude, proprietary source code, and model weights were not accessed.
- Ransom Or Extortion
- None reported
- Claim Status
- confirmed
