Cathedral AI agents reached Anthropic business systems in authorized test
Original title: Researchers Use AI Agents to Breach Anthropic’s Internal Systems in Just 12 Hours
Agents used unnamed third-party software to reach sales, HR, and IT data; prompts, source code, and model weights were not accessed.
Cathedral says AI agents, working in Anthropic’s outside-researcher program, found paths into internal business systems in under 12 hours through weaknesses in unnamed third-party administrative and sales software. Anthropic told The Washington Post the access covered a limited set of sales, HR, and IT systems holding some customer and partner contacts, not Claude prompts, source code, or model weights. Monitoring flagged some activity; access ended the same day and the software was restored only after vendors said fixes were ready. Anthropic reported no evidence anyone outside Cathedral used the weaknesses.
Unconfirmed: Anthropic and Cathedral disagree on full bug-bounty disclosure; affected vendors and flaws were not named.
- Reached data included meeting transcripts, salaries, job offers, and Slack discussion of advanced models.
- Anthropic said Cathedral did not disclose all in-system activity under bug bounty rules.
- Cathedral said it kept data access to the minimum needed for research.
- Vendor names and the exact flaws were withheld.
- Reuters reported Cathedral raised $160 million at a $1.4 billion valuation.
Impact: Limited business contacts, HR records, and internal Slack accessed
- 1.Map every connected application and limit its permissions.
- 2.Review OAuth tokens and remove access that is no longer needed.
- 3.Monitor unusual account behavior on connected business systems.
From the original article
3 min read at cybersecuritynews.com
Cybersecurity startup Cathedral says its AI agents found routes into Anthropic’s internal business systems in less than 12 hours, exposing how fast automated security tools can find and chain flaws across connected services. The work took place through Anthropic’s program for outside researchers, making this an authorized security test rather than a criminal attack, according […]
By Guru Baran · cybersecuritynews.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.