Cybersecurity startup Cathedral says its AI agents breach Anthropic internal business systems in under 12 hours during an authorised test under Anthropic’s bug bounty programme. The researchers reached meeting transcripts, employee salary data and internal Slack discussions about advanced AI models. Anthropic has disabled the affected tools and confirmed what was in scope.
The probe did not break Claude or exploit Anthropic’s own code. After several days of trying to crack Anthropic’s software directly, Cathedral shifted to third-party applications the company uses for administrative and sales work—routes that proved easier to probe.
How the AI agents got in
According to Cathedral’s draft report, the AI agents mapped possible access paths in under 12 hours. They scanned systems, tested routes and reviewed results with far less idle time for human operators. People still set the test’s boundaries and controlled how findings were disclosed.
Anthropic’s own monitoring spotted some of Cathedral’s activity before the startup formally reported it. The company cut access the same day and switched off the affected business software until vendors confirmed fixes were ready. Media assessments and the firms’ own descriptions cast the exercise as an authorised security test, not a criminal break-in.
What Anthropic says was accessed
Anthropic says the researchers reached a limited set of sales, HR and IT systems. Those held business contact details for some customers and partners, plus basic user account records. Claude customer prompts, model responses, source code and model weights were not touched, according to the company.
That distinction matters for anyone in Pakistan who uses Claude for work or study: the breach, as Anthropic describes it, sat in vendor-side business tools rather than in the model or chat layer. Still, the episode shows how an AI agents cyber attack path can open through everyday software that sits beside a core product.
Cathedral’s background and the dispute
Cathedral’s four co-founders are former staffers from the Department of Government Efficiency (DOGE). Co-founder Gavin Kliger previously served as the Pentagon’s chief data officer. In July 2026 the startup raised $160 million at a $1.4 billion valuation from Andreessen Horowitz and Sequoia Capital. Three US federal agencies have already contracted with Cathedral; a senior Trump administration official said the government is working with the firm on hostile AI scenarios.
Anthropic has accused Cathedral of overstating the national security implications of the findings. The company also said Cathedral did not fully disclose all of its activity inside the systems. Cathedral presents the AI agents breach Anthropic test as evidence that agent-driven probing can move faster than traditional human-only methods—while humans remain in charge of scope and disclosure.
For Pakistani readers following Cathedral AI cybersecurity and global AI safety, the practical lesson is familiar: the softest door is often a vendor tool, not the flagship model. Firms that rely on Claude or similar services still need to watch the third-party apps that handle payroll, sales and admin data.
Aasan alfaaz mein
Cathedral naam ki company ne AI agents se Anthropic ke office systems mein 12 ghanton se kam waqt mein raasta nikaal liya. Yeh bug bounty ke under ikazaz test tha; Claude model aur user chats safe rahe. Salary files aur meeting notes jaise business data tak access hua. Ordinary Pakistani Claude users ke liye seedha matlab yeh hai ke chat data touch nahi hua, lekin companies ko vendor software ki security pe bhi nazar rakhni chahiye.
