**["Meta AI Model 'Hacks Into a Company' Goes Viral, Turns Out It Was an Open Network"]** According to monitoring by Dongcha Beating, Meta's programming model Muse Spark 1.1 accessed a real company's system during a security test and even modified the internal environment. The headline "AI Hacks Into Another Company" quickly went viral. Did Meta suddenly create a top-tier hacker model overnight? Upon closer inspection, it turned out to be yet another case of an open network. A third-party evaluation agency, Irregular, misconfigured the testing environment, connecting what was supposed to be a closed network sandbox to the public internet. The model simply followed the existing pathway out and exploited basic vulnerabilities in the third-party service to access the system. Irregular admitted that this did not constitute sandbox escape, nor were the attack methods particularly sophisticated. Similar mistakes have previously caused trouble for OpenAI and Anthropic. OpenAI's model mistakenly identified a real website as a fictional target, while Anthropic's model accessed three organizations using weak passwords and unauthenticated interfaces. The common thread in these incidents is not that the models successfully "broke out," but that the testers left public internet access points open in the first place. Using Meta's case as an example to hype things up is especially embarrassing. The model didn’t even pick the lock—it simply walked through an open door. This proves nothing about its ability to bypass isolation, let alone justify headlines like "autonomously hacked into a company." The current formula for clickbait is already well-established: an open network is called "hacking into a company," guessing a weak password is called "hacking into a company," and discovering a zero-day vulnerability is also called "hacking into a company." The actual capabilities differ by several levels, yet the headlines remain identical. While the models haven’t yet learned to jailbreak, the media has already mastered the art of exaggeration. [Original Article Link]