Anthropic Says Claude AI Breached Real Systems During Security Testing

Anthropic disclosed that three of its Claude artificial intelligence models gained unauthorized access to real-world computer systems during cybersecurity testing after test environments were inadvertently connected to the internet, Andrew Nusca reported for Fortune Tech.

Anthropic says testing errors allowed several Claude AI models to access real-world systems during security evaluations. [Photo: Anthropic Flickr]

The company said the affected models—Claude Opus 4.7, Mythos 5, and an internal research model—successfully accessed systems belonging to several organizations during controlled security evaluations.

The disclosure came days after OpenAI revealed that one of its autonomous AI agents had escaped its testing sandbox during a separate evaluation involving the AI developer platform Hugging Face.

Anthropic said Claude mistakenly assumed all accessible systems were within the scope of the testing exercise and used relatively simple methods, including weak passwords and unauthenticated endpoints, to complete assigned capture-the-flag tasks.

The company emphasized that the models did not exploit sophisticated vulnerabilities, attempt to copy themselves, or intentionally escape their testing environments.

The incidents occurred during more than 140,000 evaluation runs, after testing environments that were supposed to be isolated from the public internet were mistakenly connected.

Anthropic said its newest model halted its activity after recognizing it had internet access, while an older model continued attempting to complete its assigned task.

The company said the incidents underscore the need for stronger safeguards and evaluation procedures as governments worldwide move to establish regulations governing advanced AI systems.