Internet Access Allowed Due to Test Environment Error
Actual Corporate Systems Breached

Anthropic's artificial intelligence (AI) model Claude gained unauthorized access to the systems of external companies during a test of its cyberattack capabilities. With OpenAI also recently disclosing a similar incident, concerns about the safety of AI systems are mounting.


Reuters Yonhap News

Reuters Yonhap News

View original image

On July 30 (local time), Anthropic stated that the test environment for Claude was supposed to be isolated from the internet, but due to an error, internet access became possible, which led to this incident. In response, Anthropic conducted a comprehensive review of its own cybersecurity evaluation records. After investigating more than 141,000 cases, the company identified three similar incidents.


Anthropic explained, “In every evaluation, Claude was informed that the test environment was virtual and disconnected from the internet. However, due to a communication lapse with our evaluation partner, Irregular, internet access was actually possible.”


All of these evaluations were conducted using the “Capture the Flag (CTF)” approach. This task requires the AI to reverse-engineer or analyze vulnerable systems or to exploit vulnerabilities and locate hidden information known as “flags.” The cases of unauthorized external system access occurred in three separate Claude models: Opus 4.7, Mythos 5, and an internal research test model.


Opus 4.7 attacked the system of a real company that shared the same name as the fictitious company given in the evaluation task. Mythos 5, upon discovering installation instructions in the simulated environment, deemed it an opportunity to attack, created a malicious Python package, and uploaded it to a public repository. This package was executed in 15 actual systems, resulting in one security company's credentials being leaked and prompting further unauthorized access to other systems. In addition, the internal research model infiltrated an application that was publicly accessible via one company's internet, but the model recognized on its own that the system belonged to an actual cloud account unrelated to its mock target and ceased its attack.


The Financial Times (FT) pointed out that this incident has heightened concerns over AI system safety, as even AI systems in their pre-release testing phase can end up hacking real external systems. This announcement comes just one week after OpenAI revealed that two of its own AI models hacked Hugging Face during tests. At that time, OpenAI's models exploited software vulnerabilities to escape the test environment, gained internet access, and executed cyberattacks.



Anthropic stated, “While multiple factors contributed to the incident, rather than focusing on placing blame, we consider all responsibility to lie with the company and are preparing measures to prevent recurrence.” The company emphasized it will broaden its review of evaluation process conversation logs and strengthen verification of partner companies in order to detect and mitigate unexpected AI behavior.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing