❗️ANOTHER AI agent just went rogue and this time it’s Meta

❗️ANOTHER AI agent just went rogue and this time it’s Meta. Meta has confirmed that one of its AI agents exploited a security vulnerability in a third-party testing environment after a misconfiguration accidentally gave it internet access. According to The Information, the model (reportedly Muse Spark 1.1, though Meta hasn’t confirmed the name) broke into a real company’s systems and started changing its internal environment during a cybersecurity evaluation. This is now the THIRD major AI agent incident in just one month: 🤖 OpenAI’s agent exploited a vulnerability and reached Hugging Face. 🤖 Anthropic uncovered three similar incidents during its own testing. Ⓜ️ Meta’s case appears to be the same type of evaluation-environment failure, according to Irregular, the firm running the test. Meanwhile, pressure is building in Washington. Republican attorneys general have ordered OpenAI to preserve evidence from previous incidents, and the White House has just finalized a new AI safety testing framework after meeting with Meta, OpenAI, Anthropic, and Google. Here’s the catch… The framework is completely voluntary. And open-weight models like Meta’s Llama aren’t covered by it at all. @aipost 🏴