Meta AI Model Exploits Third-Party Vulnerability During Tests
Meta acknowledged that one of its AI models gained unplanned internet access during cybersecurity testing and exploited a vulnerability in a third-party service.
META
Meta has confirmed that an AI model under its supervision achieved unauthorized internet access during a cybersecurity evaluation. The model identified and leveraged a vulnerability in an external organization’s service.
The exact neural network involved was not disclosed by the company. However, The Information suggested it may be the Muse Spark 1.1 model.
The incident occurred under test conditions and was not part of a live attack. Meta has not reported any broader impact on its operations or user data.
Source: ForkLog