Startup Cathedral, founded by former specialists of the US DOGE service, has demonstrated that autonomous AI systems can breach the corporate defenses of a leading artificial intelligence developer in a matter of hours. According to The Washington Post, cited by RBC-Ukraine, Cathedral researchers gained access to confidential internal information belonging to Anthropic — the creator of the Claude chatbot — as part of an official bug bounty program. The entire penetration process took less than 12 hours.
Tactics: From Direct Assault to Bypassing via Third-Party Services
Initially, the Cathedral team spent several days trying to break directly through Anthropic's secured software, but this route proved to be a dead end. The researchers then changed tactics, shifting their focus to third-party corporate services used by the company for sales management, HR tasks, and administration. By deploying autonomous AI agents, the specialists unlocked access to sensitive internal information, including corporate correspondence, in just 12 hours. This specific scenario — attacking peripheral services that integrate corporate networks rather than the security "front end" — has sparked the greatest concern among cybersecurity experts.
Limited yet Alarming Scope of Access
Anthropic emphasizes that the researchers gained access only to a restricted set of HR, sales, and IT systems. Hackers did not obtain direct access to the source code, model weight coefficients, or chatbot user queries and responses. Nevertheless, the very fact that autonomous AI agents managed to bypass the multi-layered defenses of a company working on technologies compared by Washington in potential power to nuclear weapons within a single workday has prompted a serious reassessment of corporate cybersecurity approaches across the industry.
Washington Responds: A National Security Issue
The incident provoked a wave of outrage within the administration of US President Donald Trump. A high-ranking official, speaking on condition of anonymity, stated that the Cathedral report "raises national security concerns." The official stressed that if a young startup could breach the defenses of a company developing advanced AI technologies in a matter of hours, foreign intelligence hackers could do the same with comparable ease. Cathedral, valued by investors at $1.4 billion, works closely with the US government and holds contracts with several federal agencies, adding further political resonance to the incident.
Systemic Industry Problem and Prior Warnings
Cybersecurity experts point out that the Cathedral case demonstrates not an isolated error, but a systemic problem across the entire industry. Even top AI developers remain vulnerable due to their reliance on standard third-party software that unifies corporate networks and creates single points of entry for attackers. This situation unfolds against the backdrop of lingering tension between Anthropic and the Pentagon: the company previously restricted the military use of its products over its refusal to develop autonomous weapons. Furthermore, back in September, one of the developers publicly left Anthropic, stating that the company was taking unjustified risks by relying on ordinary corporate messengers like Slack to discuss critical developments.
Anthropic's Response and Measures Taken
Anthropic representatives assured that their monitoring systems captured part of Cathedral's activities even before the official report was submitted, and access was blocked on the same day. The problematic third-party software was temporarily disconnected and brought back online only after the developers of the respective services confirmed the complete elimination of the discovered vulnerabilities. The company stated that the incident prompted a review of the architecture for interacting with external services and an enhancement of controls around the corporate network perimeter.
