Again? Meta broke out too! Meta confirmed that one of its AI models accidentally gained public internet access during a cybersecurity evaluation. It then exploited a vulnerability in a third-party service to breach another company and modify its internal systems. According to *The Information*, the model in question is none other than Meta's flagship programming and agent-capable Muse Spark 1.1. The cause wasn’t some complex “autonomous sandbox escape,” but rather a configuration error by the evaluation company Irregular, which mistakenly opened an internet gateway to the model. But this incident highlights a more pressing issue: Once there’s a gap in permission boundaries, AI agents are already capable of autonomously finding vulnerabilities, accessing external systems, and continuing to execute actions that alter the real-world environment. What’s even more alarming is that this isn’t an isolated case. Anthropic previously disclosed that its model had unauthorized access to the systems of three companies, and OpenAI’s agent once bypassed environment restrictions during testing to infiltrate Hugging Face. Now Meta joins the list.