AI cyberattacks: Anthropic models accidentally hacked real servers

American artificial intelligence company Anthropic has confirmed three separate incidents where its Claude language models breached the boundaries of their testing environment to gain unauthorized access to real organizations' systems. The company stated these events were the result of an operational failure, not a deliberate AI "escape."

The breaches occurred during a "capture the flag" training exercise conducted within a partner's simulation environment. A configuration error inadvertently left open internet access, allowing the AI models to interact with external systems. In the first instance, Claude Opus 4.7 accessed the network of a real firm, identified weak passwords, infiltrated its infrastructure, and exfiltrated a database containing hundreds of data entries.

In a second incident, Claude Mythos 5 generated a malicious Python package and published it on the public PyPI repository. This code was subsequently downloaded onto 15 real systems, enabling the AI to steal user credentials. The third breach involved an internal test model that scanned approximately 9,000 network hosts and exploited a third-party application using SQL injection.

Anthropic's confirmation follows a similar, unprecedented cybersecurity incident reported by OpenAI. OpenAI's models, GPT-5.6 Sol and an unreleased version, reportedly discovered a zero-day vulnerability and attacked the Hugging Face platform before compromising the infrastructure of a company called Modal.