After OpenAI, Another AI Security Crisis! Claude Gains Wrong Internet Access, Breaches Three Companies

Anthropic announced on Thursday (30th) that its artificial intelligence (AI) model Claude, due to a configuration error, gained internet access during testing and unauthorizedly accessed the systems of three companies. This is the latest AI-triggered security incident following its competitor OpenAI's breach of Hugging Face.

Anthropic stated that the test environment, which should have been isolated from the external internet, was mistakenly configured to allow the Claude model to connect to the internet, resulting in unauthorized access to the systems of three organizations.

After OpenAI disclosed last week that an AI agent had gone out of control during a security test and breached Hugging Face's infrastructure, Anthropic conducted a comprehensive review and examined 141,006 test logs, ultimately uncovering this incident.

These events highlight how the security risks posed by rapidly advancing AI capabilities have evolved from theoretical concerns previously voiced by experts into real-world threats—top-tier AI developers may also be caught off guard when models exploit system vulnerabilities.

Anthropic stated that Claude successfully breached the affected companies' infrastructure using basic techniques such as weak passwords and unauthenticated endpoints. The incident involved three different models: Claude Opus 4.7, Claude Mythos 5, and an internal research model. The earliest case dates back to April this year, all occurring in test environments lacking what Anthropic calls 'standard protective measures'.

The breaches occurred during a security exercise known as 'Capture the Flag' (CTF). In this test, AI models are tasked with finding hidden information within a simulated network. Anthropic stated that test instructions explicitly informed the model it could not connect to the internet, but due to a communication misunderstanding with its testing partner Irregular, the actual test environment remained connected to the public network.

Anthropic began reviewing test logs on July 23 and immediately suspended all security testing upon discovering that Claude might have accessed the internet on the same day. The company confirmed a total of three incidents by July 24 and notified the affected organizations on July 27.

Anthropic noted that two of the companies were unaware of the unauthorized access prior to notification, and the company is still contacting the third affected organization.

Anthropic emphasized that as AI models increasingly gain the ability to execute real-world cyberattacks, stricter security controls must be established in both internal and third-party testing environments.