Model safety vs. scale heats up

Two announcements landed within about 24 hours of each other: Anthropic said on April 8 that its unreleased Claude Mythos 2 Preview is being used in a new security effort called Project Glasswing, and Meta said on April 8 that Muse Spark is the first model from its rebuilt Meta Superintelligence Labs. (anthropic.com) (techcrunch.com) Anthropic’s pitch was not “our chatbot is nicer.” Anthropic said Mythos Preview has already found “thousands of high-severity vulnerabilities,” including flaws in major operating systems and web browsers, and it is putting that model into the hands of partners trying to patch real software. (anthropic.com) A software vulnerability is a hidden crack in code, like a bad lock inside a bank door. Project Glasswing is built around using a model to hunt for those cracks before criminals do. (anthropic.com) The partner list shows who Anthropic wants to impress. Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks are all named as launch partners, and Anthropic said more than 40 additional organizations that maintain critical infrastructure software also got access. (anthropic.com) Anthropic also attached money to the promise. The company said it is committing up to $100 million in usage credits for Mythos Preview and $4 million in direct donations to open-source security groups. (anthropic.com) Meta’s announcement came from the opposite direction. Instead of leading with safety, Meta led with a new model and a rebuilt machine behind it: Muse Spark is the first release from Meta Superintelligence Labs, the unit Mark Zuckerberg created after deciding Meta’s earlier artificial intelligence work was falling behind ChatGPT and Claude. (techcrunch.com) That rebuild has a price tag and a name attached to it. Meta invested $14.3 billion for a 49% stake in Scale AI and brought in former Scale chief executive Alexandr Wang to lead Meta Superintelligence Labs. (techcrunch.com) Meta says Muse Spark uses multiple artificial intelligence agents on the same problem at once, which is like putting several specialists around one whiteboard instead of asking one person to do every step alone. Meta also said Muse Spark is already available on the web and in the Meta AI app, with a coming “Contemplating” mode for harder tasks. (techcrunch.com) This fight is now centered on companies more than consumers. CNBC reported on January 21 that enterprise customers made up about 80% of Anthropic’s business and about 40% of OpenAI’s business, and OpenAI said it had more than 1 million business customers while Anthropic said it had more than 300,000 as of September. (cnbc.com) That is why these two launches look different but land on the same buyer. A bank, hospital, or large software company choosing a model provider is now comparing at least three things at once: whether the model is safe enough for sensitive work, whether the vendor can keep improving it, and whether the infrastructure behind it can survive heavy real-world use. (anthropic.com) (techcrunch.com) (cnbc.com) The tone around Anthropic has shifted in public too. On April 7, CNBC aired Big Technology founder Alex Kantrowitz saying “OpenAI is chasing Anthropic right now,” and on April 7 BFM Tech ran a segment titled “Anthropic ready to surpass OpenAI,” which is not how challengers are usually described unless buyers are already taking them seriously. (cnbc.com) (youtube.com) So the story is no longer just who has the flashiest demo. Anthropic is trying to win trust by showing a model that can defend critical systems, while Meta is trying to win confidence by showing it can rebuild fast, spend heavily, and ship a new model through a new stack. (anthropic.com) (techcrunch.com)