The Morning Build for October 16, 2026: Meta's Muse Spark Open on Mac, Figma Adds Agent Specs, and Persistent Safety Gaps

Meta released Muse Spark Open as an open-weight, multimodal model optimized to run on Apple M5 Ultra Mac Studios; coverage and hands-on reporting probe both its capabilities and safety limits. Figma moved design automation toward production with markdown Specs and public code-layer beta; independent researchers published a preprint showing leading image models remain easy to jailbreak for explicit outputs.

Meta releases Muse Spark Open weights, a multimodal 816B model with 1M-token context optimized for Apple M5 Ultra

What happened: Meta published Muse Spark Open on its research blog and released weights on Hugging Face and major clouds; the model is 816 billion parameters with 42 billion active, supports text, image, video, and audio, and advertises a 1,000,000-token context window. Meta says Muse Spark Open runs on two Apple M5 Ultra Mac Studios with 512 GB unified memory each and provides Helion GPU kernels and MLX optimizations for multiple inference engines.
Why it matters: Engineers can run a high-context, agent-capable open-weight model locally on Apple silicon, or deploy it via hosted partners and major cloud model registries; license terms allow free use for many developers but require a label and restrict API commercial providers with over $20 million in revenue to a separate agreement.
Outlook: A full evaluation report and the Muse Spark Open model card will be published soon, per Meta, and those documents will be the next authoritative source for the model’s safety evaluation thresholds and measured risk categories.

Sources: the-decoder.com · techcrunch.com

Meta lists compatibility and hosting partners for Muse Spark Open: Hugging Face, SageMaker, Foundry Models, and inference engines

What happened: Meta stated Muse Spark Open weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, and that the model runs on inference stacks including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Meta named hosted inference partners including Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter.
Why it matters: Developers have immediate options to run the open-weight model locally, on third-party inference engines, or via hosted partners, enabling experimentation across hardware (AMD, Apple, Intel, NVIDIA) and deployment modes without waiting for a single vendor runtime.
Outlook: Availability across multiple clouds and inference engines sets up short-term interoperability checks and independent performance comparisons once third-party benchmarks and user reports appear in the coming weeks.

Sources: research.meta.ai

WIRED reports Muse’s consumer-facing agent engages in explicit sexting during real-world tests

What happened: WIRED’s Uncanny Valley segment and article describe reporter tests where Meta’s Muse agent produced sexually explicit content when prompted; WIRED reports the interactions were straightforward to elicit and quotes Meta as framing Muse as for adults 18 and over and allowing, quote, ‘legal adult behavior.’
Why it matters: First-party agent behavior on consumer channels provides an operational view of guardrail effectiveness at scale, showing that a deployed agent with broad capabilities can generate explicit outputs despite platform-level content moderation tensions across Meta products.
Outlook: Meta’s public statements and the forthcoming Muse Spark Open evaluation materials are the next concrete sources to reconcile the agent’s deployed behavior with the safety metrics Meta says it applied during release.

Sources: wired.com

Figma opens code layers and an agent-oriented Markdown Specs editor in public beta to steer AI in design workflows

What happened: Figma announced a Markdown-based Figma Specs editor that can reference canvas elements, embed FigJam diagrams, code layers, and integrate design and code libraries; the company also made code layers available in public open beta and said agents can consult centralized Specs and libraries to follow brand and design rules. Figma added support for over 20 tools to its MCP, and integrated the Payload headless CMS into Figma Make.
Why it matters: Design-to-code workflows now expose a machine-readable specification and code-layer editing inside the canvas, giving engineers and product teams a concrete artifact agents can read and act on to reduce mismatch between design output and production code.
Outlook: Figma’s public open beta for code layers is the immediate test window where enterprise and individual users will generate feedback and bug reports that will determine how agent-driven design-to-code flows behave in production prototypes.

Sources: techcrunch.com

White Circle preprint shows DALL·E 3, Nano Banana 2, and GPT-Image: 2 can be bypassed to produce explicit images

What happened: Ars Technica summarized a White Circle preprint and press release reporting that using off-the-shelf tools and API calls the researchers bypassed safety filters on DALL·E 3, Google’s Nano Banana 2, and GPT-Image: 2; measured explicit-image hit rates were about 13 percent for DALL·E 3, under 10 percent for Nano Banana 2, and about 3 percent for GPT-Image: 2 in their tests.
Why it matters: The results quantify practical bypass success rates against deployed image generators and show that even modern models with policy layers can return explicit outputs at nontrivial rates when subjected to automated attack methods, increasing the need for engineering-level mitigations at model, API, and hosting layers.
Outlook: White Circle’s preprint and press materials from last month are the recent public disclosure; vendor responses, security patches, or updated red-team results from the affected model providers will be the next concrete signals to watch.

Sources: arstechnica.com · arstechnica.com · arstechnica.com