Anthropic AI Models Hack Three Organizations During Cybersecurity Testing

Anthropic’s Claude AI models successfully breached the computer systems of three separate organizations during a series of cybersecurity evaluations, the company reported.

The disclosure, detailed by The New York Times, CNBC, and Politico, comes as part of an internal effort to identify vulnerabilities and risks inherent in the company’s artificial intelligence systems. It marks only the second time a major AI developer has admitted that its models successfully infiltrated outside infrastructure during a testing phase.

Controlled Breaches and “Real-World Incidents”

The incursions were not accidents. According to a company statement, Anthropic conducted “real-world incidents” investigations to determine if its models could facilitate cyberattacks. The models were specifically tasked with identifying and exploiting security flaws within a controlled evaluation process.

Reporting from The Washington Post confirms the models broke into three organizations. Anthropic has not publicly named the affected entities.

Beyond Simple Text Generation

This is not a matter of mere autocomplete. By navigating the security perimeters of three different organizations, the Claude models demonstrated a level of reasoning and tool-use that exceeds basic text generation.

The models identified vulnerabilities and executed exploits without requiring direct human intervention for every step. This capacity for autonomous hacking tasks is exactly what Anthropic is attempting to neutralize. The company stated the goal of these tests is to build “guardrails” to prevent the AI from being used for unauthorized access or data theft in public or commercial settings.

The Risk of Offensive Cyber Capabilities

Anthropic now joins a small group of AI labs admitting their models possess offensive cyber capabilities. For regulators and security researchers, the concern is clear: LLMs could lower the barrier to entry for sophisticated attacks, enabling non-experts to launch targeted breaches.

The company has not yet detailed the specific methods used to penetrate the systems or the nature of the data accessed. However, Anthropic maintains these evaluations are critical for developing safer systems, asserting the breaches resulted from intentional testing rather than a failure of safety filters.

The Practical Reality of “Dual-Use” AI

The disclosure arrives as pressure mounts on AI firms to provide “red teaming” reports—simulated attacks designed to find weaknesses. These results provide concrete evidence of the “dual-use” nature of AI, where a tool built for coding assistance is repurposed for exploitation.

Cybersecurity experts have long warned that as models become more proficient at writing and debugging code, their ability to find “zero-day” vulnerabilities—previously unknown software flaws—will increase. Anthropic’s results confirm that this theoretical risk is now a practical reality.