AI startup uses autonomous agents to breach Anthropic systems

The breach of Anthropic

The AI startup Cathedral recently demonstrated how quickly autonomous agents can compromise a major tech company. After human researchers failed to break into Anthropic's systems for several days, the startup deployed AI agents that succeeded in less than 12 hours. As reported by TechRadar, the breach was conducted as part of a bug bounty program intended to find and fix security flaws.

The agents managed to infiltrate Anthropic's network by identifying and exploiting vulnerabilities in third-party business software. While Anthropic has strong internal defenses, the incident revealed that external vendors can serve as a weak link. The startup eventually shared its findings with Anthropic, though the company later accused them of failing to disclose the full extent of their activity.

Anthropic confirmed that the breach resulted in the exposure of a limited set of data. This included information from human resources, sales, and IT systems, as well as contact information for certain customers and partners. However, the company stated that its source code and AI model weights remained secure throughout the incident.

What the agents accessed

Once the agents gained entry to the network, they were able to view a variety of internal communications and sensitive documents. This included meeting transcripts, Slack messages, and specific human resources data such as salaries and job offers. The breadth of the access shows how much internal information is often stored within standard business tools.

Notably, the breach did not appear to expose employee conversations with Claude, Anthropic's primary AI model. The focus of the data access seemed to be on organizational and administrative records rather than the core AI interactions. This distinction is important for understanding what types of data are most at risk during a third-party software compromise.

The startup, Cathedral, is led by former members of the US Department of Government Efficiency (DOGE). The founders have backgrounds in the Pentagon, SpaceX, and investment banking. Their goal is to provide offensive and defensive cyber operations for the US military and government, making their technical capabilities significant.

The role of third-party risk

The incident serves as a major wake-up call regarding the security of ordinary enterprise software. Cathedral noted that weaknesses in common tools can compromise the infrastructure of even the most advanced organizations. This is a classic example of a supply chain attack, where a hacker targets a less secure vendor to reach a larger, more secure target.

Cybersecurity experts suggest that security maturity is often limited by the weakest vendor a company integrates with. Even if a company like Anthropic uses world-class engineering for its own products, it is still vulnerable to the flaws in the software it buys from others. This means that no amount of internal hardening can fully protect a company if its partners have poor security practices.

Anthropic's response was relatively quick, with the company stating that its monitoring flagged the activity before the startup reported it. They were able to shut off the agents' access on the same day the breach was detected. Despite this quick reaction, the fact that the breach occurred at all highlights a persistent problem in modern corporate IT environments.

Implications for the industry

The timing of this breach is particularly sensitive for Anthropic as the company prepares for its Initial Public Offering (IPO). With a valuation of $1.4 billion, the company needs to maintain a high level of trust with investors and partners. Leaking sensitive data, even to a white-hat group like Cathedral, can create a negative signal regarding the company's security posture.

The incident also highlights the growing power of AI-driven cyberattacks. While human hackers may take days or weeks to find a way into a system, autonomous agents can perform these tasks in a fraction of the time. This shift requires companies to move toward more proactive and automated defense mechanisms to keep up with the speed of AI-powered threats.

Moving forward, the industry will likely see a greater focus on independent verification of security patches. Experts argue that simply closing a known entry point is not enough; companies must ensure that no other