From BigGo Finance · BigGo Finance · · 1m 50s
AI Agent Breaches Anthropic's Network in 12 Hours
AI agent from startup Cathedral breaches Anthropic's internal network in 12 hours, exposing payroll records and Slack chats.
In brief
AI agent from startup Cathedral breaches Anthropic's internal network in 12 hours, exposing payroll records and Slack chats. Cathedral's AI agent accessed Anthropic's internal systems in under 12 hours during a security test, revealing payroll and Slack data without leaking core AI models. Originally reported by BigGo…
Security Test Breach
U.S. AI company Anthropic's internal network was breached in under 12 hours by an AI agent from startup Cathedral during an authorized security test.
Leaked Content Details
The breach accessed meeting transcripts, compensation details, hiring information, and internal slack messages but not claude user conversations, core source code, or model weights.
Cathedral's Background
Cathedral, founded by former U.S. Department of Government Efficiency members, has a $1.4 billion valuation and secured contracts with three federal agencies.
Breach Method
The AI agent exploited security vulnerabilities in third-party commercial software to access Anthropic's sales, HR, and IT systems without targeting core defenses.
Anthropic's Response
Anthropic confirmed the intrusion but stressed that claude user conversations, core source code, and model weights were not leaked and accused cathedral of exaggerating national security risks.
Security Concerns
The incident raised concerns in Washington about frontier ai assets and third-party software risks, with a senior trump administration official calling it a material threat.
Industry Impact
Similar breaches have occurred recently, including an openai employee arrested for placing a recording device and researchers taking over chatgpt accounts using claude.
Market Shifts
The incident may accelerate changes in enterprise security spending and ai security testing, with vendors using ai agents to conduct automated penetration tests winning federal contracts.
Next Steps
For anthropic and similar labs, the real test is how to redraw an enforceable boundary between third-party software ecosystems and core asset protection.
At a glance
Cathedral's AI agent accessed Anthropic's internal systems in under 12 hours during a security test, revealing payroll and Slack data without leaking core AI models.
