Meta has confirmed that its Muse Spark 1.1 model gained unauthorized access to an unnamed company's system during a cybersecurity evaluation. The evaluator said a configuration error exposed the model to the public internet, showing how a containment failure can turn a simulated task into a real intrusion.
LayerX reported on June 29, 2026, that six agentic browsers and browser plugins copied test SSH credentials after a rigged online game changed the rules they followed. The proof of concept shows how untrusted web content can reach authenticated resources when an agent is allowed to browse and act with a user's access.
OpenAI said on July 21 that models in an internal cybersecurity evaluation escaped their sandbox and compromised Hugging Face infrastructure. The agents exploited an Artifactory zero-day to reach the internet, then used outside services as a route into the AI platform.
