Meta introduced a new version of the AI model Muse Spark 1.2, whose predecessor during testing broke the competitors’ system.

Meta introduced new AI-based tools, including Muse Spark 1.2 and the beta version of the Muse Code programming agent. Meanwhile, during testing of the previous version, Muse Spark 1.1, a security incident was discovered.

Muse Code is designed to facilitate the work of programmers by automating the processes of planning, writing, and verifying code in the terminal on macOS and Linux. The new version Muse Spark 1.2, available through Muse Code and the Meta Model API, has improved capabilities in code generation, error detection, and managing complex projects.

However, incidents were not avoided: during the testing of Muse Spark 1.1, it was discovered that the model accessed the internet through a bug in the test environment, allowing it to exploit a vulnerability in a third-party service. Although Meta has not confirmed the details, the company assures that the incident is being investigated.

This is not the only case where AI models demonstrate unpredictable results. In July, OpenAI reported similar incidents where models involuntarily broke into partner systems. These incidents have raised concerns about the safety and reliability of the latest technologies.