Good morning. It's Thursday, April 9, and we're covering Anthropic’s locked-down Mythos model, its new agent architecture, Meta’s latest push into multimodal AI, and more.

If you enjoy this email, share it with a friend. First time reading? Sign up here to keep up with the future of tech.

Meta Unveils Muse Spark AI Model
Meta introduced Muse Spark, a multimodal AI model with tool use and multi-agent reasoning. It supports tasks like coding, visual analysis, and health insights. A new “Contemplating mode” uses parallel agents to boost performance. The model is available now, with API access in limited preview.

Anthropic Withholds New Model Over Cyberattack Risk
Anthropic made the rare decision to keep Mythos Preview out of public hands — and safety testing results explain why. The model appeared to know when it was being evaluated, deliberately underperformed on one test to seem less suspicious, and in one experiment emailed a researcher from an isolated machine with no supposed internet access. (more on this below)

Tubi Launches ChatGPT Streaming App
Tubi launched a native app inside ChatGPT, enabling users to discover content via natural-language prompts. The app connects users to Tubi’s 300,000+ titles with personalized recommendations. It marks the first major streamer to build directly within ChatGPT rather than its own platform.

AI Predicts Heart Failure Early
University of Oxford developed an AI tool that predicts heart failure risk up to five years in advance. Trained on 72,000 patients, it achieved 86% accuracy using routine CT scans. The model detects hidden inflammation signals around the heart. High-risk patients were 20x more likely to develop the condition. Researchers are seeking approval for clinical rollout.

Recraft V4 introduces two powerful model families — Raster for high-resolution photorealistic scenes or detailed illustrations and Vector for scalable SVG graphics with design taste at the core of every output. Available inside Recraft Studio, Recraft V4 supports complex prompts, crisp typography, and professional-grade outputs across branding, packaging, and digital workflows.

Start creating with Recraft V4.

Anthropic has developed a new AI model, Claude Mythos, capable of uncovering thousands of previously undetected software vulnerabilities—including flaws dating back 27 years—and is choosing not to release it publicly. Announced on April 7, 2026, the model’s capabilities raised concerns that it could accelerate cyberattacks if misused.

Instead, Anthropic is sharing Mythos with select partners like CrowdStrike, Palo Alto Networks, Amazon, Apple, and Microsoft under a defensive initiative called “Glasswing.” The move highlights a growing tension in AI: tools that strengthen cybersecurity can also empower attackers at unprecedented speed and scale. → Read the full article here.

Anthropic describes a new architecture for long-running AI systems called “Managed Agents,” designed to separate core components—model, execution environment, and memory—to improve reliability and adaptability. The system decouples the “brain” (Claude model and harness), “hands” (tools and sandboxes), and “session” (event log), allowing each to scale or fail independently.

This shift reduced latency significantly, with time-to-first-token dropping ~60% at median levels and over 90% at the high end. The approach also strengthens security by isolating credentials from AI-generated code and supports more flexible deployments, including enterprise environments. → Read the full paper here.

Engineers at the University of Southern California have developed a memristor-based memory device that operates at temperatures up to 700°C (1,300°F), far beyond the ~200°C limit of conventional electronics. Published on March 26, 2026 in Science, the device uses a tungsten–hafnium oxide–graphene structure to prevent heat-induced failure and maintain performance.

It can store data for over 50 hours, endure more than one billion cycles, and compute at low voltage with nanosecond speeds. The breakthrough could enable AI processing in extreme environments like space, geothermal systems, and nuclear settings, while significantly improving energy efficiency for matrix-heavy workloads. → Read the full article here.

World Labs announced two updates to its Marble model family, aimed at improving visual quality and scaling capabilities for AI-generated environments. Marble 1.1 enhances lighting, contrast, and reduces visual artifacts, while Marble 1.1-Plus is designed to generate larger, more complex virtual worlds.

The models are available via a user interface and API, signaling a push toward both creator-friendly tools and developer integration. The release reflects growing competition in generative 3D and simulation platforms. → Read the full article here.

OpenAI Unveils Child Safety Blueprint: OpenAI introduced a U.S.-focused framework targeting AI-enabled child exploitation, emphasizing updated laws and built-in safeguards developed with NCMEC and safety groups.

Apple Quietly Opens Door to NVIDIA: Apple approved a third-party driver enabling NVIDIA GPUs on Apple Silicon Macs in limited external setups, signaling cautious flexibility as AI workloads increasingly demand NVIDIA.

Bezos AI Lab Poaches Key Talent (Paywall): Jeff Bezos’s Project Prometheus hired Kyle Kosic from OpenAI to build industrial-focused AI systems, intensifying competition for talent and expanding into real-world intelligence.

Google Finance Expands with AI Globally: Google is rolling out its AI-powered Finance platform to 100+ countries, adding natural-language research, advanced charts, real-time data, and live insights with multilingual support.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team