AI agents breach Anthropic systems in under 12 hours

The breach of Anthropic

Cathedral, a cybersecurity startup, recently demonstrated the speed of modern AI-driven hacking by infiltrating Anthropic's internal systems. According to the Washington Post, the company's AI agents managed to gain entry into the systems in less than 12 hours. The breach was conducted as part of a program modeled on a bug bounty, which is an arrangement where outside experts are paid to find and report security flaws.

The researchers were able to access a limited set of internal systems including HR, sales, and IT departments. These systems contained sensitive business information such as meeting transcripts, salary details, job offers, and Slack messages regarding advanced models. However, Anthropic confirmed that the researchers did not touch any customer conversations with Claude, nor did they access the company's source code or model weights.

Anthropic stated that their internal monitoring systems flagged the activity before the startup officially reported it. The company acted quickly to shut off the startup's access on the same day the breach occurred. They also took the step of pulling the specific vulnerable tools out of service until the third-party vendors could provide signed repairs.

Vulnerabilities in the supply chain

The method of entry provides a clear look at how modern corporate networks are often interconnected. Rather than attacking Anthropic's core software directly, the AI agents took advantage of flaws in outside vendors' admin and sales tools. Many large companies rely on these same types of third-party tools to manage their daily business operations, making them a common target.

The draft report from Cathedral suggests that the agents were able to navigate these external tools to reach internal data. This highlights a significant risk in the software supply chain, where a single vulnerability in a minor tool can provide a gateway to a larger organization. It shows that even high-profile AI companies can be exposed through the software they use to manage their employees and sales.

Anthropic has noted that the researchers made several days of failed attempts to get into their software head-on before finding a way in. This suggests that the perimeter of the main software was robust, but the

The founders and the funding

The startup behind the breach is led by four former Department of Government Efficiency staffers. The founders of Cathedral are Gavin Kliger, Luke Farritor, Marko Elez, and Jack Stein. Their team focuses on cybersecurity with a specific concentration on military applications for their technology.

Despite being a new company, the startup has already secured significant financial backing from investors. According to Decrypt, the group closed a funding round of $160 million in July. This investment resulted in a company valuation of $1.4 billion, highlighting the massive capital flowing into AI-driven security tools.

Rule violations and security fears

The relationship between the two companies was complicated by reports of rule-breaking and exaggerated claims. Anthropic stated that Cathedral did not fully follow the rules of the research program because some of its activities were never disclosed. Furthermore, the company noted that the startup exaggerated the national security implications of the breach.

Despite these criticisms, the incident sparked significant alarm within the government. A senior Trump administration official noted that the report has been circulated among government agencies, prompting serious concerns about national security. The official expressed worry that if a new startup could achieve this, other nations like China might be able to do the same.

Scope of the data access

The specific systems accessed by the researchers were limited to those managing sales, HR, and IT functions. These internal databases contained business contact details and basic account records for various users within the organization. Anthropic confirmed that the breach did not extend to the core components of their AI models or private customer interactions.

The company also stated that they have found no evidence to suggest that any other group or individual exploited these specific vulnerabilities. This indicates that the security flaw was isolated to the activity monitored during the research period. By identifying these specific points of entry, the company can focus its remediation efforts on the third-party tools involved.

National security and military focus

Cathedral is a cybersecurity startup that utilizes AI technology with a primary focus on military applications. This specialized concentration on defense-related security gives the startup a unique profile in the tech industry. Their ability to navigate complex systems quickly has drawn attention to the capabilities of new AI-driven security firms.

The speed of the breach has prompted high-level government concern regarding international competition. A senior official noted that if a new startup could achieve such a breach, other nations like China might be able to do the same. This sentiment highlights a growing fear that small, agile AI companies could pose significant risks to national infrastructure.