Muse Code: Meta's Answer to AI Coding Agents

Mark Zuckerberg just entered the AI agent race with Meta's own coding agent. Meta just launched its new coding agent this week called Muse Code. It's built on a model called Muse Spark 1.2. It spins up multiple sub-agents that can write, test, and fix your code in parallel. So, instead of one AI agent writing all your code, Muse uses a team of AI agents that can do things simultaneously. During a cybersecurity evaluation, one of the models from the Muse Spark family got internet access it wasn't supposed to have. [music] And once it had that access, it found vulnerabilities in an unrelated company's systems and exploited them all on its own. Now, Meta says this wasn't the AI breaking out of its sandbox. [music] Instead, it was a setup mistake by the outside company who was handling the test. But, either way, the AI found a small opening and used it to hack a very real company, and nobody told it to do this. And this isn't even the first time this has happened this year. Other labs have had similar incidents >> [music] >> during their own safety tests. This is what moving fast can look like if the AI agent that you're building can act on itself. Would you trust an AI agent with server access? Tell us in the comments below, and for everything AI, subscribe to Pleebo AI.