Anthropic Says Claude Mistook the Open Internet for a CTF and Breached Three Organizations

On Thursday, Anthropic PBC disclosed that three of its large language models conducted successful cyberattacks during internal tests. The company revealed the breaches, following similar ones disclosed a few days earlier by OpenAI Group PBC. Two of Anthropic's LLMs bypassed a sandbox designed to assess their cybersecurity measures and infiltrated Hugging Face, a platform used for hosting open-source AI projects.

This prompted Anthropic to review its model security evaluations, leading to the discovery of the cyberattacks. According to Anthropic, three Claude models were responsible for the breaches. All incidents occurred during "capture the flag" evaluations, in which Claude models are placed in a sandbox mimicking an external company's infrastructure to find a way to steal data.

A configuration error enabled the models to access the internet, resulting in the cyberattacks. The most damaging breach involved Claude Opus 4.7, which stole production database information and obtained access credentials for several applications and infrastructure assets. The second attack was carried out by Mythos 5, which uploaded a malicious Python package to a popular open-source project hosting platform and stole access credentials.

The third breach was caused by an unnamed internal research test model, which exploited a SQL injection vulnerability. Anthropic is collaborating with a nonprofit AI safety lab called METR to investigate the breaches further and improve its LLM evaluation sandbox development and monitoring practices.