# Meta's finance chief jumps to Fireworks as fal raises $1.3B > Fireworks hires Meta veteran Dave Wehner as CFO, fal banks $1.3B at a $16B valuation, Ant ships Ling-3.2-flash, and Apiiro flags leaked secrets in Claude Code repos. Source: https://ai.dosa.dev/blog/today-in-ai-2026-11-08 Published: 2026-11-08 Author: Dosa AI Tools --- This brief covers AI news from 2026-11-08 UTC. Money moved at the top of the stack today, while a security shop found credentials sitting in public repos after a Claude Code permission prompt. ## Fireworks hires Meta veteran Dave Wehner as CFO **What happened:** Dave Wehner, who spent eight years as Meta's finance chief, is leaving the company to become CFO of Fireworks. Fireworks was valued at $17.5 billion in July and is in early talks about another funding round. **Why it matters:** A CFO hire and fresh round talk at an inference provider usually precede pricing and capacity changes for teams serving open models through it. [Source: AI Affairs]() ## Fal raises $1.3B at a $16B valuation led by Bessemer **What happened:** Fal, a San Francisco startup running video, image and audio models, raised $1.3 billion at a $16 billion valuation. Bessemer Venture Partners led, with Nvidia, Sound Ventures and T. Rowe Price joining. **Why it matters:** More capital for a media-model inference vendor means builders get another well-funded endpoint to route image and video generation through. [Source: Axios]() ## Ant Group ships Ling-3.2-flash with a 1M-token window **What happened:** Ant Group released Ling-3.2-flash, a native multimodal MoE with 560B total and 25B active parameters, a 1M-token context and text, image and video input. Chat and OpenAI/Anthropic-compatible APIs are live and free for two weeks; weights stay closed. **Why it matters:** Compatible endpoints plus a two-week free window make it cheap to test long-context multimodal calls without rewriting client code. [Source: Neural Daily]() ## Apiiro: Claude Code prompt left passwords in 1, 122 public files **What happened:** Application-security firm Apiiro says in a blog post that Claude Code's "Don't Ask Again" permission choice exposed passwords in 1, 122 public GitHub files. Anthropic says the file is meant to stay local. **Why it matters:** If your agents can write outside the repo, audit which permission prompts your team has silenced before the next push. [Source: AI Tracker]() ## OpenCode Browse lets an agent drive your real browser **What happened:** OpenCode said on November 6 that OpenCode Browse is available. It lets an agent control the user's own browser, with their open tabs, logins and extensions, working through MCP and a Chrome extension. **Why it matters:** Agents inheriting live sessions and logins remove the auth scaffolding from browser automation, and remove the sandbox with it. [Source: AI Tracker]() ## NVIDIA posts four open Cosmos 3 Nano world models **What happened:** NVIDIA released four 16B-parameter Cosmos 3 Nano open-weight checkpoints for interactive world simulation, covering multi-sensor driving data, streaming observations, simulator-to-RGB transfer and bimanual robot action-to-video. **Why it matters:** Open weights at 16B let robotics and simulation teams fine-tune a world model locally instead of licensing one. [Source: Neural Daily]() ## Scale Labs tests coding agents on drug discovery tasks **What happened:** Scale Labs ran Claude Code (Opus 4.7), Codex (GPT-5.5) and Gemini CLI (Gemini 3.1 Pro) on 66 expert-curated biomedical tasks in a shared tool environment. GPT-5.5 led structural reasoning, Opus 4.7 led patent mining, and the longest pipelines defeated all three. **Why it matters:** Long-horizon tool chains, not single answers, are where coding agents still break, which is where your orchestration and retries matter most. [Source: Scale Labs]() ## Sources - [AI Affairs]() - [Axios]() - [Neural Daily]() - [AI Tracker]() - [Scale Labs]()