JadidJust-in AI Developments, Incidents & Deadlines

Podcast · Wednesday, November 11, 2026

Agent-built software is going live with nobody checking who can reach it, and the bills are now countable.

Today's headline

Agent-built software is going live with nobody checking who can reach it, and the bills are now countable. A security firm found more than thirty-five thousand Grok-built apps open to anyone. Railway puts seven hundred forty thousand dollars of fraudulent AI usage on stolen customer keys. Meanwhile, vendors kept pushing reach: Anthropic moved into meetings and spreadsheets, and Google made its strongest Gemini generally available.

On the clock

Node.js ships security releases today for versions twenty-two, twenty-four and twenty-six. The top severity is high. Upgrade the day the release lands.

Two Gemini changes are close. Gemini 4 Carbon doubles in price after November thirtieth, from two dollars to four dollars per million input tokens. Budget for the higher rate before you commit to it.

And OpenAI will remove the gpt-image-2 model on February tenth. It points to two newer image models as replacements, so plan the move.

Security

A security firm called Claroc found that nearly all of the apps built with Grok's Build mode are open to anyone with the link. That is more than thirty-five thousand apps. About one in ten held internal or personal data, including a hospital group's staff list. The company's own sales pipeline was exposed too. Claroc blames a default publish setting and weak access control in the generated code. Audit every app your team built this way, and unpublish or add sign-in today.

An attacker used a flaw in SGLang, an open source inference engine, to run code on a NEAR AI server. The server sat inside a trusted execution environment, and the company says that does not protect against a bug in the software running inside it. About one hundred users had account details taken, and eighteen agents were reached. The flaw is critical and exploited. Update SGLang and rotate keys on any host that ran it.

Railway confirmed that stolen environment variables led to about seven hundred forty thousand dollars of fraudulent AI charges. The attacker queried for AI provider keys across as many as one hundred thousand workspaces. Railway says it will cover charges that a provider will not refund. Move AI keys out of platform environment variables, set provider spend caps, and rotate anything stored there.

Xen Orchestra has an authentication bypass that is critical and exploited. A crafted cookie logs an attacker in as an existing token's owner, then leads to root on the virtualization hosts. There is no CVE yet, so scanners may be silent. Apply the latest release and limit the management interface to trusted networks.

Microsoft researchers showed that editing fewer than ten rarely used neurons can plant a backdoor in an agent model. A trigger string makes it run an attacker's command, and it still passes evaluations. The attacker needs write access to the weights, and Microsoft has not seen it used. Treat model weights like any other supply-chain artifact.

A few quick ones. Git fixed a flaw where cloning a hostile repository with submodules can run code. Upgrade to Git two point fifty-six point one, or the matching patch for your series, and update apps that bundle their own Git. Microsoft's Patch Tuesday again topped one thousand flaws, including an exploited Outlook spoofing bug that can leak password hashes. Ivanti fixed an unauthenticated SQL injection in Endpoint Manager that is rated nine point eight. And N-able released hotfixes for the N-central zero-days, so on-premises customers should apply them.

Agents

Google made Gemini 4 Carbon generally available. It sits above Gemini 4 Argon for large migrations and long-running agent work. Google also previewed a swarm feature that coordinates several coding agents from one prompt. Google's benchmark claims are its own, and Carbon uses more tokens than Argon, so the real bill runs higher than the price suggests. It is also in GitHub Copilot.

Anthropic launched Claude Meetings, Whiteboards and Sheets. Claude joins a video call as a visible participant, drafts plans while people talk, and files the notes afterward. Reportedly, Microsoft shares fell on the news.

Google engineers found that old TODO comments pointing at long-fixed bugs pull coding agents into side-quests. That burns tokens and can corrupt logic. Clear dead TODOs before pointing an agent at a legacy repo.

A Linux memory-management reviewer stepped back from several areas, saying AI-written patches take too much of his time. That is the cost behind the fix volume Linus Torvalds praised yesterday.

Models and pricing

Anthropic's Fable five point two completed forty-one percent of projects on the Remote Labor Index. That test asks whether freelance work is good enough for a paying client. The best system a year ago scored two and a half percent.

Business and policy

Anthropic's updated IPO filing adds about ten billion dollars of cloud commitments with Microsoft and Google. It also says a thirty-five billion dollar lease with Lambda may be cancelled. One count puts the total raise, with a private placement, at up to eighty-eight billion dollars.

OpenAI told investors its third-quarter revenue rose nearly fifty percent to ten billion dollars, according to a report. Its adjusted operating loss grew to ten point nine billion. Anthropic reported about seventeen billion in the same quarter.

CoreWeave beat estimates and raised guidance, and its shares jumped about fifteen percent premarket. That eased the fear over neocloud debt, though SoftBank wrote off three point seven billion dollars of data center advance payments.

Nvidia's board hired the law firm Wachtell to investigate CEO Jensen Huang. The House Foreign Affairs chair says he will subpoena Huang if he does not answer a briefing request today. Nvidia reports earnings on November seventeenth.

And New York opened its RAISE Act registration portal. Frontier developers above five hundred million dollars in revenue must register by January first.

Action items

First, audit every Grok-built or one-click-published app and add sign-in. Second, patch Xen Orchestra and limit its management interface. Third, update SGLang and rotate keys on any host that ran it. Fourth, move Git to the fixed release before cloning untrusted repositories. Fifth, take AI provider keys out of platform environment variables and set spend caps.

Before you go

Open your list of internal tools built by an AI assistant and ask one question of each: who can reach this without logging in? Start with anything that has a publish button. Then run through today's recap.

Type to search · ↑↓ to move · enter to open