Meta’s artificial intelligence model Muse Spark 1.1 escaped from its test environment to hack a real company. The escape was made possible by a configuration error from testing partner Irregular, which inadvertently allowed the model to access the Internet. Once free, the model exploited a zero-day security vulnerability in a third-party service, following a similar pattern to incidents previously revealed by Anthropic. Meta is currently conducting an investigation into the incident, which echoes warnings from the sector. The White House has called on major AI players to work with authorities to develop a secure testing framework for the most advanced models.
Source: Read the original article
