AI News Roundup: Anthropic Mythos Leak, OpenAI Kills Sora, Meta Bets on RISC-V

Anthropic’s Mythos Leak Rattles Cybersecurity Markets

The biggest story in AI this week centers on a model nobody was supposed to know about yet. Claude Mythos (codenamed Capybara), accidentally exposed on March 26 through an unsecured content management system, represents what Anthropic calls a “step change” in AI capability — a new tier above Opus that scores dramatically higher on coding, reasoning, and cybersecurity benchmarks.

Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered the exposed data store, which contained a draft blog post describing Mythos as “currently far ahead of any other AI model in cyber capabilities.” Internal safety documents warn the model could “presage an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”

Markets reacted swiftly: CrowdStrike dropped 7%, Palo Alto Networks declined 6%, and Zscaler fell 4.5% on Friday. Anthropic says the exposure was a CMS misconfiguration, not a breach, and that it will focus initial Mythos deployment on defensive cybersecurity applications. A formal announcement is expected to be accelerated.

OpenAI Shuts Down Sora After Burning $1 Million Per Day

Just six months after its splashy consumer launch, OpenAI announced on March 24 that it is shutting down Sora — both the mobile app (closing April 26) and the API (closing September 24). The AI video generation tool peaked at roughly one million users before collapsing to under 500,000, all while consuming approximately $1 million per day in GPU costs.

The numbers tell the story: Sora generated roughly $2.1 million in total revenue while burning through hundreds of millions in compute. The Disney licensing partnership that allowed characters inside Sora is also winding down. TechCrunch called it “a reality check moment for AI video” as OpenAI pivots resources toward its planned IPO and core language model products.

Meta Unveils Four RISC-V AI Chip Generations

Meta revealed an aggressive custom silicon roadmap in mid-March: four generations of MTIA chips — the 300, 400, 450, and 500 — all built on the open-source RISC-V instruction set architecture. The company chose RISC-V over Arm to avoid licensing dependencies, with TSMC manufacturing and Broadcom co-development.

The performance trajectory is staggering: a 25x increase in compute FLOPs and 4.5x increase in HBM bandwidth from the MTIA 300 to the 500. Meta plans a new chip generation every six months, with the 300 already in production for ad ranking and recommendations. The 400 through 500 will handle generative AI inference through 2027, powering image generation, video synthesis, and the recommendation engines behind Meta’s ad business.

Shopify Brings Commerce to Every AI Chat

Starting March 24, Shopify’s Agentic Storefronts went live for all eligible U.S. merchants — enabled by default with no setup required. Products from millions of Shopify stores are now discoverable and purchasable through ChatGPT, Google AI Mode, Microsoft Copilot, and the Gemini app, with orders flowing directly into the Shopify admin.

This marks a pivot from OpenAI’s earlier “Instant Checkout” approach: purchases now complete on the merchant’s own storefront rather than inside the AI conversation. Shopify reports AI-driven traffic to its stores is up 7x since January 2025. A new “Agentic” plan also lets brands not currently on Shopify add products to its catalog and sell across AI channels.

White House AI Framework Targets State Law Preemption

The administration’s National Policy Framework for Artificial Intelligence, released March 20, is the most detailed federal AI legislative blueprint to date. Its central thrust: replace the growing patchwork of state AI laws with unified federal rules, arguing that fragmented regulation stifles innovation.

Key proposals include no new federal AI regulator (existing agencies keep sector-specific oversight), mandatory child safety controls, standardized data center permitting, and — most controversially — blocking states from regulating AI model development or imposing liability on developers for third-party misuse. The White House wants legislation “this year.”

Eli Lilly Powers Up Pharma’s Most Powerful AI Supercomputer

Eli Lilly inaugurated LillyPod, the pharmaceutical industry’s most powerful AI supercomputer, equipped with 1,016 NVIDIA Blackwell Ultra GPUs delivering over 9,000 petaflops of AI performance. Assembled in just four months at Lilly’s Indianapolis campus, the system can simulate billions of molecular hypotheses in parallel and aims to cut the traditional 10-year drug development timeline in half.

NVIDIA and Eli Lilly also announced a first-of-its-kind AI co-innovation lab, with up to $1 billion in combined investment over five years spanning talent, infrastructure, and compute for drug discovery.