AI and IT News Recap: August 11, 2026: Meta Shrinks a Frontier AI Onto Your Laptop, a Perfect-10 Metabase Zero-Day Loots Real Companies, and Claude Code Goes Hands-Off

By Noah Smith, Owner & Consultant, KeyChange Technologies · August 11, 2026

Since our last recap on Friday, the biggest AI and IT news has been about power moving closer to home: Meta put a frontier-class agent on a laptop, Anthropic let its coding agent off the leash, and attackers turned a perfect-10 analytics bug into real customer data theft. Here is what actually matters for your business today.

📌 The AI and IT news at a glance

• 🔝 Meta's Muse Glimmer puts a capable AI agent on a single laptop and gives it away under an open license.
• 🤖 Anthropic makes Claude Code "auto mode" the default, so the agent stops asking permission at every step.
• 🍽️ ChatGPT can now book your dinner table through Yelp, OpenTable, and Resy without leaving the chat.
• 🛡️ A perfect-10 Metabase zero-day was exploited in the wild, and real companies confirmed stolen customer data.
• 🔒 New StormEncryptor ransomware from a China-linked crew is riding a remote-management flaw into managed networks.
• 🧩 A WordPress supply-chain hack at BdThemes quietly minted rogue admin accounts across popular Elementor plugins.

🤖 AI

🔝 Meta shrinks a frontier-class agent down to your laptop

Meta released Muse Glimmer, a 30-billion-parameter open-weights model built for "always-on" local agent work: multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery, all in a single model that runs on your own machine with no cloud connection required. The headline trick is size. Running a 30B model at full precision normally needs more than 55GB of memory, but a 4-bit quantized build drops it under 20GB, so it fits on a single consumer GPU or a well-specced laptop with 24GB or 32GB of memory. Meta shipped it under an Apache 2.0 license with day-one support for Ollama, LM Studio, llama.cpp, MLX, vLLM, and more, and it is already downloadable from Hugging Face. It is a distilled cousin of Meta's larger Muse Spark model, positioned as the "personal intelligence" tier of Zuckerberg's roadmap.

This is a different animal from the Muse Code coding model we covered on August 6. Muse Glimmer is about running a general-purpose agent entirely on hardware you already own, which changes the math on privacy, latency, and cost for anyone nervous about sending data to a vendor's API.

In short: Meta released Muse Glimmer, a 30B open-weights agentic model that runs locally on a single consumer GPU under an Apache 2.0 license.

What it means for your business: A private, offline AI agent that lives on a laptop is now realistic, which is a genuine option for teams handling sensitive data or working where cloud access is unreliable. It also means "we can't use AI because of data residency" is a weaker excuse than it was last week.

My take: The interesting shift here is not benchmark scores, it is the address the model lives at. When a capable agent runs on a laptop with no network, the entire conversation about data leaving your building changes. I would not throw out your cloud models, but I would start a small pilot on genuinely sensitive workflows, because local-first is going to keep getting better and cheaper.

Source: Meta releases Muse Glimmer, a 30B open agentic AI model that runs locally on PCs (Neowin)

🤖 Anthropic makes Claude Code stop asking permission

Anthropic announced that "auto mode" will become the default in Claude Code for Pro, Max, and Team accounts starting August 14. In auto mode, the coding agent proceeds through its work without pausing for human approval at each step, unless an action is judged to be irreversible, destructive, or aimed outside your own environment, in which case it still stops to check. It is a small toggle with a big philosophical message: the default posture for agentic coding tools is shifting from "ask first" to "act, then flag the risky stuff."

For teams already living in Claude Code, this mostly means fewer clicks and faster loops. For everyone else, it is a useful signal of where the whole category is heading, and a prompt to think about the guardrails you want before the agents get more autonomous by default.

In short: Starting August 14, Claude Code runs in auto mode by default for paid tiers, acting without step-by-step approval except on irreversible or destructive actions.

What it means for your business: Agentic tools are quietly moving the default from "confirm everything" to "confirm only the dangerous things," so it is worth deciding now what your team considers dangerous, and making sure those actions are the ones that trigger a human check.

My take: Defaults are policy. Most people never change them, so "auto mode on by default" is Anthropic making a bet that the agent is trustworthy enough to drive. I think that is roughly right for day-to-day coding, but the value of this change lives entirely in that "irreversible, destructive, or outside your environment" carve-out. If you adopt it, spend your attention there, not on the speed.

Source: Anthropic is turning Claude Code's auto mode on by default (TechCrunch)

🍽️ ChatGPT will now book your dinner reservation

OpenAI and Yelp expanded their partnership so ChatGPT can find and book restaurant tables inside the conversation. Tell it where and when you want to eat, your party size, and preferences like cuisine, budget, or dietary needs, and it will surface options and complete the booking. Reservations run globally through OpenTable, in the US through Resy, and in the US and Canada through Yelp, including waitlist joins. It is rolling out across ChatGPT plans on mobile, web, and desktop, and it builds on Yelp licensing its reviews and business data to OpenAI in late July.

In short: As of August 10, ChatGPT can book restaurant tables and join waitlists directly in-chat via Yelp, OpenTable, and Resy.

What it means for your business: This is another brick in "agentic commerce," where the assistant completes the transaction instead of just handing you a link. If you run a business that takes bookings or appointments, the discovery surface is quietly moving into chat assistants, and being present there is starting to matter.

My take: Booking dinner is the demo, but the real story is transactions happening inside the assistant. The consumer version is dinner tables; the business version is your customers asking an AI to schedule, buy, or reschedule without ever touching your website. Worth watching which booking platforms these assistants choose to integrate, because that is where the customers will flow.

Source: Yelp brings restaurant reservations and waitlists to ChatGPT (Search Engine Land)

🛡️ IT and security

🔝 A perfect-10 Metabase zero-day is stealing real customer data

Metabase, the widely used open-source business-intelligence and dashboard tool, warned that a maximum-severity flaw (CVSS 10.0) was exploited in the wild as a zero-day. The bug sits in the password-reset endpoint and lets an unauthenticated attacker inject arbitrary SQL, escalate to administrator access, steal the credentials for every database Metabase connects to, and read or export the data behind them. Metabase Cloud instances have been patched automatically, but self-hosted deployments need to update by hand, and patched releases are out across the supported branches.

This one already has named victims. Hardware maker Framework and finance tool Tally both confirmed that customer information was accessed after their Metabase instances were compromised, including names, email addresses, login IP addresses, and billing and shipping details. Because Metabase typically holds the keys to your other databases, security researchers have described the potential blast radius as unusually wide.

In short: A CVSS 10.0 unauthenticated SQL-injection zero-day in Metabase was exploited in the wild, and Framework and Tally confirmed stolen customer data.

What it means for your business: If you or a vendor runs self-hosted Metabase, treat this as urgent: patch now and assume connected-database credentials may be exposed. Even if you do not use Metabase, it is a reminder that your analytics layer often holds credentials to everything else.

My take: A perfect-10 in a tool that, by design, stores logins to all your other databases is close to a worst case. The patch is the easy part. The harder and more important part is rotating every credential that Metabase could see, because if attackers got in before you patched, the door they walked through is your database passwords, not the Metabase bug itself.

Source: Metabase Zero-Day Exploited in Wild Allows Admin Access Without Authentication (The Hacker News)

🔒 A new ransomware crew is pivoting through remote-management tools

Microsoft flagged a new ransomware strain called StormEncryptor, deployed by Storm-1175, a China-linked, financially motivated group previously tied to the Medusa ransomware operation. StormEncryptor is C++ malware that appends a ".encrypted" extension to files and drops a ransom note demanding payment within three days or the stolen data leaks. The recent intrusions were likely preceded by exploitation of an authentication-bypass flaw (CVE-2026-18577) in N-able's N-central remote monitoring and management platform, the kind of tool managed service providers use to reach across many client networks at once.

The pattern is the part to notice. Attackers increasingly break into the management console that sits above dozens of businesses, then use its built-in remote-control features to push ransomware down into every endpoint it touches.

In short: China-linked Storm-1175 has begun deploying new StormEncryptor ransomware, likely entering through an actively exploited N-able N-central flaw used by managed service providers.

What it means for your business: If an outside IT provider manages your systems, their remote-management tools are now a top target, and a compromise there lands directly on you. Ask your provider what they run, whether it is patched, and what monitoring sits around it.

My take: The uncomfortable truth is that the software your IT partner uses to help you is also the fastest way for an attacker to reach all of you at once. This is not a reason to distrust managed providers, it is a reason to ask them harder questions. A provider who can answer clearly about patch timelines and account controls on their RMM is worth keeping.

Source: New StormEncryptor ransomware used by former Medusa affiliate (BleepingComputer)

🧩 A WordPress plugin vendor got hijacked without changing a line of code

A supply-chain compromise hit BdThemes, a popular vendor of WordPress add-ons for the Elementor page builder, and it worked in an unusually sneaky way. Instead of tampering with plugin source code in the official WordPress repository, attackers poisoned a static JSON data feed that BdThemes plugins fetch for an internal promotional banner system. A parsing flaw introduced back in March turned that feed into a cross-site-scripting vector, letting attackers hijack administrator sessions, create rogue admin accounts, and install persistent web shells on affected sites, all without any plugin update being released. Wordfence flagged the campaign, and the impacted plugins, including Element Pack, Prime Slider, and Ultimate Post Kit, were pulled from the repository pending investigation.

In short: Attackers poisoned a JSON feed used by BdThemes WordPress plugins to silently create rogue admin accounts and deploy web shells, with no source-code change or update required.

What it means for your business: If your site runs BdThemes or Elementor add-ons, check for unfamiliar admin accounts and unexpected files now, not later. More broadly, it shows that "we did not install any updates" is no longer proof you were not compromised.

My take: This is the supply-chain attack that breaks people's mental model. Everyone has learned to be careful about updates, but here the malicious payload arrived through a background data feed the plugin quietly trusted. The lesson for any website owner is that trust flows through every external thing your software phones home to, not just the code you consciously install.

Source: BdThemes plugins supply-chain hack creates rogue WordPress admins (BleepingComputer)

Want the previous edition? Read our AI and IT News Recap for August 7, 2026.