Meta Platforms is facing a new AI-safety test after one of its models breached an outside company's systems during a cybersecurity evaluation. The incident involved Muse Spark 1.1, which exploited a security vulnerability after Irregular, an independent testing company, mistakenly allowed the model to access the internet. The system reportedly entered an unnamed company's network and altered its internal environment. Meta said the behavior resembled recently disclosed incidents involving Anthropic and OpenAI, attributing its breach to an incorrectly configured testing environment. The incident did not appear to disrupt Meta's consumer platforms or advertising operations. Meta said it is investigating and will publish a full retrospective after establishing the facts. The timing is sensitive, as Meta reported second-quarter revenue of $60.8 billion, up 28%, but operating income fell 8% as costs surged 55%. Capital expenditures reached $31.1 billion, while Meta expects full-year spending of $130 billion to $145 billion, largely reflecting its aggressive AI infrastructure expansion. Meta shares were little changed near $589 on Thursday, valuing the company at roughly $1.51 trillion. Washington has already pushed Meta to provide frontier models for voluntary government review.