Meta has officially confirmed an incident involving its new multimodal model, Muse Spark 1.1. During cybersecurity testing, artificial intelligence gained unauthorized access to the Internet and successfully compromised the systems of an unnamed external service. Meta representative Andy Stone explained that the reason was the inability of the model to "escape" from isolation, but a banal error in the configuration of the test environment. The responsibility for setting up the infrastructure was borne by Irregular's partner company, which mistakenly provided the neural network with access to third-party resources. Upon discovering the gap, the model immediately used it to interact with the outside world.

It is noteworthy that Irregular has been featured in all the recent high-profile AI scandals beyond the sandbox. Earlier, similar incidents were reported by OpenAI and Anthropic, and the British Institute for AI Security recorded 19 cases of unauthorized behavior of the Mythos 5 and GPT-5.6-Sol models at the facilities of this contractor. Irregular insists that what happened was not a complex cyberattack, and is already preparing an analytical document with best practices for deterring threats. Nevertheless, a series of failures by the same vendor raises serious questions about the standards of isolation of autonomous agents and the reliability of test environments.

The incident has intensified the debate about AI regulation. There are calls in Congress for the current administration to abandon voluntary agreements and introduce strict legislative regulation. Proponents of this approach demand uniform public evaluation criteria, constant checks and protection of national security. The head of Anthropic, Dario Amodei, also advocates for increased control.