October 22, 2026 — GitLab's perfect-ten bug and the upgrade nobody wants

Host0:32

GitLab patched seven release branches on Wednesday, and one bug in them scores a perfect ten point oh. The flaw is an integer overflow in the regular-expression compiler. That just means a number outgrew its container. Hand the compiler a specially crafted regular expression, and under certain conditions, an unauthenticated attacker could execute arbitrary code. The Hacker News says the engine is Onigmo, Ruby's regex engine, shipped in GitLab's packages. The affected range reaches back to GitLab seven point oh. GitLab.com is already patched. If you use GitLab.com, you're covered. If you self-host, this is your afternoon. The fixed releases run from nineteen point five point one down to eighteen point eleven point thirteen, for both CE and EE. In GitLab's words: “we strongly recommend that all self-managed GitLab installations be upgraded to one of these versions immediately.” Now the catch. GitLab hasn't disclosed what conditions let an attacker reach the vulnerable compiler. The Hacker News reports there's no workaround, so the patch is the only certain protection. And that ten is GitLab's own advisory score, not an independently reproduced exploit. Second catch, downtime. Upgrading to nineteen point four point three or nineteen point three point six runs database migrations. The Hacker News reports that means downtime on single-node installs. GitLab says its multi-node zero-downtime procedure avoids that. GitLab also published two detection rules, Onigmo native worker crash and Org mode render error burst. Wire them into alerts while you schedule the upgrade. The release fixed eleven security issues in all. And GitLab backported fixes to branches older than its normal window of the current and two previous monthly releases. That one regex spans seven branches. Until someone explains how a request reaches that compiler, you can't rule yourself out. So check your version, find your branch's fixed patch, and follow the update instructions in GitLab's notice. Patching is the part of the job nobody applauds. Here's something more fun to download, a free coding model its makers say was largely built by an AI.

3:02

Flint has forty billion parameters, and only about two and a half billion of them wake up for any given token. That's the pitch for a coding model Recursive says runs on a laptop. It's open-weight under Apache two point oh, released Wednesday. Weights are on Hugging Face, and OpenRouter hosts it free behind an OpenAI-compatible API. The model ID there is recursive slash flint dash v one point zero, colon free. Ollama's zero point forty point eight release supports it too, on macOS first, with other GPUs to follow. Context is two hundred sixty two thousand, one hundred forty four tokens. You get tool calls and JSON response formatting, but not JSON-schema enforcement. LM Studio says to keep the reasoning content Flint returns in your conversation history. Free endpoints are rate-limited, and no number is published. Under the hood, it's a sparse mixture of experts, so most of the model sits idle on each token. LM Studio describes forty nine layers, thirty five with sliding-window attention and fourteen with full attention. Recursive says its Eureka Machine helped with nearly every step, a claim that isn't independently verified. Now the numbers, all reported by Recursive and listed by LM Studio. Seventy four point two percent on SWE-bench Verified, using mini-swe-agent. Fifty four point nine percent on Terminal-Bench two point one, using the Terminus two harness. Recursive also measured about one hundred thirty one tokens per second on an M four Pro MacBook Pro, against fifty eight to sixty nine for comparable models. Company-run, not independent. Fine print. Laptop-sized comes with a twenty four gigabyte minimum, and LM Studio warns that may not fit the full context. MarkTechPost says the launch blog cited up to five hundred twelve thousand tokens. The published pages say two hundred sixty two thousand, so trust those. My verdict. Try it on a small repo investigation, a bug fix, or a test-writing task, and guard tool execution. Don't drop a proven assistant over vendor-reported scores, especially for big autonomous jobs. A coding model can write the code that renders a video. Hearing what that code made is a different problem, and today's repo goes after it.

5:39

In the README's words, AI coding agents can now write the code that renders a film, a product demo or a game. They still cannot hear what they made. Earshot is the attempt to fix that. It's open-source, CPU-only, MIT licensed, and exposed as a CLI, a Python API, and an MCP server. You point it at audio or video, and it hands back descriptions, loudness, timing, sync, and spectrum measurements. Under the hood, it pairs CLAP text-audio embeddings and an AudioSet tagger with objective measurements. Its forge makes sounds from physical-synthesis recipes, scores candidates against a description, and uses evolutionary search to mutate candidates. Version zero point three adds two hundred one effects stored as code, no models or network needed. The number that matters, project-reported: on a held-out test, a strong sound-kind match was right ninety eight percent of the time. The exact variant was right sixty one percent. The README is candid about limits. Its search optimizes its own models, which can be fooled, so it says to check spectrograms and run human blind tests. It can't judge taste, emotion, provenance or licensing, and one blind test used a single listener. The twenty two second launch film uses sounds made entirely in code, no recordings and no music track. To try it, git clone the repo, create a virtual environment, and install the CPU PyTorch build and Earshot extras. First use downloads two models, about one point one gigabytes. Earshot lets an agent check its own work. Anthropic's Duet is an agent that watches your apps while you're away.

7:32

Anthropic's sample morning brief gives its new agent a very human errand. Flag a six hundred forty dollar county health-permit fee, and ask you to approve payment before it expires. That's Claude Duet, a research preview powered by Opus five point five. It's a waitlist for Claude Pro and Max subscribers in the United States, on the web or in a standalone iOS app, according to MacRumors. The interface is one continuing thread. You message it. It draws on your account's memory, skills and connectors, works on its own cloud computer, and returns finished documents or asks you to decide. Anthropic says it works with more than six thousand connectors and plugins. It asks approval before sensitive steps like sending messages or making payments. MacRumors adds that it checks your apps every few hours, uses a virtual machine with its own browser, and runs a dedicated classifier. Anthropic's own page warns it may misread a thread, misjudge urgency, or miss something that lands between check-ins. Background checks count against your plan's usage limits, and Team and Enterprise plans aren't included yet. Notice what this isn't. It's a consumer preview, not a generally available developer API. Privacy-minded users and enterprise admins will balk at wiring business apps to a cloud agent. The privacy and safety claims are Anthropic's own description of its architecture. What I can't tell you is how much usage those background checks eat. Anthropic gives no fixed quota, so until someone measures real use, that's the number to watch. An agent that holds logins is only as safe as the runtime and the credentials underneath it. Wednesday's ship log has both.

9:29

Node twenty six point twelve point one and twenty four point twenty two point one. Upgrade first. Ten vulnerabilities are fixed, led by a Brotli decoder out-of-bounds write that permits remote code execution, plus high-severity DNS and TLS validation flaws. The twenty two line is covered too, and end-of-life versions are affected as well. One medium fix stopped entropy caches from being saved in startup snapshots, which had every process repeating the same random values. Vercel routing middleware now speaks Python, for FastAPI, Flask, Django and others. It runs before requests reach your app and can rewrite, redirect, or answer them. Install it with uv add, using the vercel dot proxy dependency group, then point vercel dot toml at a proxy entrypoint. It runs as a separate app, so your redirect now has its own dependency group. Vercel's example redirects old blog slugs with a three oh eight, and doesn't say how latency or limits compare with its JavaScript options. GitHub CLI two point one oh three point zero adds opt-in expiring tokens, in public preview. Run gh auth login with the expiring flag, and gh stores a short-lived access token with a rotating refresh token. GitHub says a leaked token becomes unusable within hours instead of lasting indefinitely. For HTTPS Git, also run gh auth setup-git, with Git two point forty six or newer. Anything that copies the output of gh auth token once may break when it expires, so read it fresh. All of that assumes you've got a machine to run it on. Nvidia spent Wednesday in Berlin on how fast a data center can show up.

11:23

DSX POD is Nvidia's reference design for partner-built Vera Rubin AI factories, assembled and validated inside modular containers before delivery. Nvidia says that can cut deployment at a prepared site from six months to as little as fourteen days. That's a projection, not a measured deployment, and partners are expected to deliver in early twenty twenty-seven. ServeTheHome reports seven Vera Rubin NVL72 racks per container, still needing outside power, networking and cooling. It also says Nvidia isn't building or selling these itself. The two items closer to your desk: OpenShell, a reference architecture for agentic AI safety, and a DGX Spark update that supports clusters of up to eight units. If you run agents on your own hardware, those are the two to read. Nvidia promises a data center in fourteen days. These next two stories needed far less time to cause trouble.

12:29

Fast Boot, where nothing shuts down when you ask it to. LiteLLM, poisoned. A bad LiteLLM release sat on PyPI for one hour and forty four minutes and planted a stealer that runs whenever Python starts. The lookalike repo was Berri-AI, the real one is BerriAI, so the disguise was one hyphen. Signal, still listening. A bug rated nine point three could leave your microphone live after hangup if a caller flooded the call's event queue. The Register's advice is: “watch what you say after hanging up a call,” which is a strange line for a CVE. Same theme twice: software that kept running after you said stop, a stealer at Python startup and a microphone after goodbye. If you ran the bad LiteLLM, assume compromise, pin to one point one hundred six point two, and rotate your secrets. And update Signal.

13:26

Today was GitLab's perfect ten patch, the Flint coding model, Earshot, and Claude Duet, plus Node, Vercel and GitHub fixes in the ship log. Your try-it for today is checking your GitLab version, and patching if it's behind. Links to everything are in the show notes. Follow The Daily Reboot in your podcast app, and tomorrow's episode finds you. I'm your host. Go patch something.