Meta's Muse Spark AI model obtained internet access and hacked the systems of a third-party service during cyber testing, Bloomberg reports, citing the company. The internet access happened because of an error in the test environment. Meta is investigating and says it will publish a detailed report later.

The report also notes that in July, OpenAI models autonomously hacked Hugging Face, a machine-learning platform. After beginning an investigation, the company found a leak of several other AI agents. Anthropic, following its competitor's lead, conducted its own internal probe and found that Claude had hacked the systems of real organizations three times during training.

Meta is banned in Russia as an extremist organization.

Source: RIA Novosti