AI DAILY / DEV
MONDAY
October 12, 2026

    Reflection AI Unveils Beam, a 501B Open-Weight MoE Pitched as 'DeepSeek of the West'

    • 501B-parameter sparse MoE with 23B active per inference, trained on 23.8T tokens; aimed at coding, reasoning, and agentic workloads.
    • Reflection claims it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks using 3-4x less inference compute — vendor numbers, not independently verified.
    • Apache 2.0 weights, technical report, model card, and dev tooling promised later this month; early access is waitlist only, no public API or price yet.
    • Nvidia-backed Reflection framing it as the US answer to Chinese open-weight leaders; Fireship covered it alongside Mistral Large 4 and Moonshot in his Oct 7 'open-weight race' video.
    models techcrunch.com

    Meta, Sierra, Walmart, Stripe and Shopify Launch Personal Agent Protocol

    • Oct 7: open standard for authenticating personal AI agents to businesses, with read-only or write permissions and visibility into agent activity.
    • Agents can connect as a guest; customers decide the access level each company gets, and merchants see which agent is acting on their site.
    • Founding backers include Meta, Sierra, Walmart, Stripe, Shopify, Genesys, Instinct, and Rocket; Sierra plans to publish v0.1 of the spec later in October.
    • Third parallel standard this quarter: Walmart and Shopify also co-sign Google's Universal Commerce Protocol, while Stripe helped build the Agentic Commerce Protocol with OpenAI.
    industry channelinsider.com

    Pwn2Own Ireland Debuts AI Category — LiteLLM Falls Twice, Codex and Dynamo Breached

    • Oct 6-8 contest in Cork paid $1,262,000 for 98 unique zero-days across 61 successful attempts by 29 teams.
    • AI infrastructure track hit for the first time: LiteLLM breached twice in one day, OpenAI Codex cloud agent fell to a single argument-injection bug, NVIDIA Dynamo and Chroma also compromised.
    • Taisic Yun (Xint) chained an input-validation flaw with code injection for a reverse shell on LiteLLM ($40k); HaeJung Yang (Out of Bounds) took Dynamo on day two for another $40k.
    • Ikotas Labs won Master of Pwn with $361k and 42.5 points; signals AI gateways and coding agents are now first-class attack surfaces alongside printers and phones.
    research bleepingcomputer.com

    'REA: Reverse Engineer Anything' Rockets Up GitHub Trending With 25K Stars in a Day

    • morluto/rea: +25,599 stars today, now at ~117k total — a CLI and MCP server that lets coding agents reverse engineer software through Hopper Disassembler.
    • Ships 50+ MCP tools covering function dossiers, Evidence v2 reproducible records, and controlled process capture; installs via `npx skills add morluto/rea` or the `rea-agents` npm package.
    • Roadmap extends to packaged apps, JavaScript bundles, websites, firmware, mobile artifacts, and runtime-behavior diffs between versions.
    • Matthew Berman cosigned it this week — fits the broader 'skill packs' trend that owned GitHub trending in early October.
    open-source github.com

    OpenAI Codex CLI 0.161 Makes GPT-6.1 Sol the Default Coding Model

    • Oct 7 release carries the GPT-6.1 Sol default forward across the bundled and Amazon Bedrock catalogs; model ID is `gpt-6.1-sol`.
    • Adds `/mcp login` for signing into MCP servers from the terminal, plus microphone and speaker selection for voice conversations.
    • Amazon Bedrock path picks up multi-agent V2 and Ultra reasoning on compatible models; Daybreak stays opt-in.
    • Rollout is per-account — some ChatGPT-tied users still see 'model not supported' errors; downgrading Codex doesn't fix it.
    tools github.com