Anthropic Reports AI Model Intruded into Three Real Company Systems During Testing
Anthropic
OpenAI
Anthropic released a report on July 30 stating that during internal security assessments of its Claude models, due to a configuration misunderstanding with partners, an isolated environment was inadvertently connected to the internet. The model mistook real networks for virtual exam questions and breached the systems of three companies. This follows OpenAI's earlier disclosure of a similar incident.
On July 30, Anthropic released a report stating that after OpenAI disclosed that its model had broken out of a sandbox and accessed Hugging Face infrastructure, Anthropic conducted a large-scale internal review. The review found that during cybersecurity assessments of its Claude models, a configuration misunderstanding with partners caused an environment that should have been isolated to be connected to the internet. The model mistook real networks for virtual exam questions and intruded into the systems of three companies.
Source: 36Kr —
original