Meta releases first coding agent, reports first agentic hack

Meta's superintelligence team released Muse Code in beta on Wednesday, a terminal coding agent built on its new Muse Spark 1.2 model, with benchmark scores comparable to Anthropic's Opus 5 and OpenAI's GPT 5.6 Terra models. The same day, The Information reported that Meta's Muse Spark 1.1 model accidentally gained internet access during testing and breached a third-party company's internal systems, with Meta blaming sandbox environments built by testing partner Irregular.

AI /tag/ai/ Meta’s superintelligence team is following in the footsteps of the industry’s biggest AI labs with two major milestones: Its first coding agent, and its first agent to hack a third-party company. It released Muse Code in beta https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2?ref=thestack.technology on Wednesday, a terminal coding agent built on its new Muse Spark 1.2 model, and shared benchmark scores comparable to Anthropic’s Opus 5 and OpenAI’s GPT 5.6 Terra models. On the same day, a report in The Information https://www.theinformation.com/articles/meta-ai-model-hacked-another-company-cybersecurity-testing?ref=thestack.technology revealed its Muse Spark 1.1 model had accidentally been given internet access during testing and breached a company’s internal systems. As with the similar Anthropic incident reported last week https://www.thestack.technology/anthropic-our-models-can-breach-containment-as-well/ , Meta placed the blame on sandbox environments built by testing partner Irregular that were mistakenly given internet access. Get the full story: Subscribe for free Join peers managing over $100 billion in annual IT spend and subscribe to unlock full access to The Stack’s analysis and events. Subscribe now https://www.thestack.technology/membership/ Already a member? Sign in https://www.thestack.technology/signin/