Muse Glimmer: 30B-Parameter Model for Local Agent Workflows
866 points · Meta AI Research
Meta Superintelligence Labs dropped Muse Glimmer today — a 30-billion-parameter open-source model (Apache 2.0) purpose-built for always-on local agent workflows. The pitch: run it on a Mac or PC with a single consumer GPU, do function calling, local coding, and LLM-as-a-judge evaluation without touching the cloud.
The training pipeline is interesting: logit distillation from a larger teacher model (Muse Spark), mid-training on longer-context agent-heavy data with reasoning traces, then post-training combining SFT with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. They claim strong results on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. Integrations with llama.cpp, MLX, and ExecuTorch are coming “in the coming days.”
The local-first angle is compelling for anyone who’s tired of API dependency and latency. A 30B model that actually handles tool calling reliably on consumer hardware would be a genuine step change. The real test is whether “strong performance” means actually usable in production agent loops or just benchmark-competitive — those are very different things. Apache 2.0 is a solid licensing choice, though; no strings attached for commercial use.
Docker Sandboxes: Disposable Isolated Environments for AI Agents
566 points · Docker
Docker launched Sandboxes — isolated, disposable microVM environments for AI coding agents like Claude Code, Copilot CLI, Codex, OpenCode, and Kiro. The pitch: “YOLO mode, safely.” Agents get full autonomy — install packages, modify configs, spin up their own containers — while the host stays untouched. Install via brew install docker/tap/sbx on macOS, winget install Docker.sbx on Windows.
The product is positioned as solving the trust problem with agentic coding: agents do their best work when unconstrained, but you don’t want them nuking your host filesystem. Each agent gets a dedicated microVM with only the project workspace mounted. Network and filesystem controls are configurable, and org-wide enforcement comes via Docker AI Governance (a separate enterprise product).
This is a smart move by Docker — they’re inserting themselves as the infrastructure layer for the AI agent ecosystem. The free tier hooks developers, and governance features upsell to enterprises. The real question is latency: microVM spin-up time matters when agents are iterating fast. And “disposable by default” means you’d better have good state management, or every agent restart loses context.
tl;dv Leaks 181,874 Meetings in Wide-Open Firestore Database
437 points · bobdahacker.com
tl;dv (Too Long; Didn’t View), the AI meeting recording platform with 2+ million users, has its Firestore database wide open. Security researcher BobDaHacker reported this on January 28, 2026. Six months later — still not fixed. The CTO never responded to the disclosure.
The vulnerability is embarrassingly simple: any authenticated tl;dv user gets a Firebase token that queries the meetings collection with zero tenant isolation. That’s 181,874 meeting records across 84,312 users on 35,003 email domains. Government meetings from 23 countries (including .gov domains from the US, Brazil, Ukraine, and Japan). University meetings from Berkeley, University of Tokyo. Corporate meetings from HubSpot, Confluent, and 35,000 other domains. The records include conference IDs — which are joinable Google Meet or Teams rooms. At any given moment, roughly 1,000 meetings with status “recording” have exposed, live conference IDs.
The researcher demonstrated the attack by joining a Malaysian Ministry of Education call with 157 participants and a US university startup pitch with 21 students screen-sharing their project. Six months of irresponsible non-disclosure from tl;dv makes this worse. When your entire business model is recording sensitive meetings and your security posture is “any free-tier user can query everything,” you’ve fundamentally failed at the one thing that matters.
Mars Bar from 1991 Found — 20g Bigger Than Today’s
227 points · BBC News
A 35-year-old Mars Bar was found during a house clearance in Scunthorpe, and the viral photo tells the whole story: 62.5g in 1991 versus 40g today. That’s a 36% reduction in product size over three decades.
Victoria Gordon, who runs the cleaning service Pocket Rockets, posted the comparison photo and it went nuclear on social media. Mars’s corporate response was the usual: “updates to our bar sizes and pack formats to reflect consumer demand, alongside considering wider external factors such as manufacturing costs and the price of cocoa.” Translation: we charge you more for less and call it innovation.
Gordon says she might “do a UK tour” with the bar rather than eat it. Fair enough — at this rate, a 1991 Mars Bar is worth more as a museum exhibit than as food. Shrinkflation isn’t new, but a physical artifact from 35 years ago makes the scale of the deception visceral in a way that charts and percentages never can.
The Throughline
Today’s HN front page is dominated by one theme: trust boundaries in the age of AI agents. Meta’s Muse Glimmer wants you to trust a local model with your personal context. Docker’s Sandboxes want you to trust agents with full system autonomy — just inside a box. tl;dv demonstrates what happens when you build an AI-powered platform and completely forget about tenant isolation. And the Mars Bar, while not tech, is the same story in a different domain: companies quietly eroding what you get while maintaining the illusion nothing changed.
The tl;dv story is the most alarming. 180,000+ meetings exposed for six months after responsible disclosure. Government calls, university strategy sessions, corporate sales pitches — all queryable by any free-tier user. This is what happens when the AI gold rush prioritizes shipping features over basic security architecture. Firestore with no tenant isolation is a first-year mistake, not a production system serving 2 million users.
Meta’s local model push and Docker’s sandbox play are two sides of the same coin: the industry is slowly acknowledging that sending all your data to the cloud isn’t sustainable. The question is whether local-first AI actually delivers on its privacy promises, or just moves the attack surface from the cloud to your device.