AI Tech Deck

This week in AI: the five things that mattered

Week ending Sunday, September 21, 2026. A security flaw that hit every major coding agent, Google's best voice models yet, Meta's new subscription, OpenAI's first formal incident reports, and a lawsuit over the slowdown. Each item ends with the only question we care about: does this change what you should use?

By HarryPublished September 2026

A fanned hand of five glossy cards in muted jewel tones spread on a dark navy table, lit from one side
AI-generated illustration

1. A zero-click flaw hit Claude Code, Codex, Copilot and Gemini CLI

On Thursday, September 17, researchers at Air Security published Plugin4Shell, a vulnerability in how the four most widely used AI coding agents install plugins. Marketplaces pin each plugin to a specific commit hash. The agents fetched that hash but never checked what actually landed on disk, so whoever controls a plugin's repository could create a branch named after the hash, fill it with malicious code, and have every agent's background auto-update install it. No click required, and the plugin runs with the developer's full access. Anthropic fixed it in Claude Code 2.1.179 and OpenAI in Codex 0.146.0, both before disclosure. Google told the researchers it will not patch the Gemini CLI because it is being retired, and points users to Antigravity. Microsoft had not shipped a Copilot fix; GitHub says its own hosting blocks hash-shaped branch names, though the researchers note marketplaces on Bitbucket and self-hosted git are still exposed. There is no sign the flaw was used in a real attack.

Source: The Register, September 17; Air Security's report.

Does this change what you should use? Yes, today: update Claude Code and Codex, and if you are on the Gemini CLI, move off it.

2. Gemini 3.8 Live made Google's voice mode the one to beat

On Tuesday, September 15, Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its new speech-to-speech models. The Extended Thinking model reasons and talks at the same time, narrates progress on multi-step tasks, and took the top spot on Artificial Analysis's speech-to-speech quality index with a score of 82.6. Both models handle interruptions, switch between 97 languages mid-conversation, and run tool calls in the background while the conversation continues. Availability is broad: 3.8 Live is in Search Live for everyone, and Extended Thinking is in Gemini Live for everyone, in Docs for Google AI Pro and Ultra subscribers, and in Gmail and Keep for all Google AI subscribers. All generated audio is watermarked with SynthID.

Source: Google, September 15.

Does this change what you should use? Not what you pay for, since the upgrade reached free Gemini Live, but if voice is how you use an assistant, Gemini now deserves a head-to-head with ChatGPT's GPT-Live.

3. Meta started charging for more AI inside Instagram, WhatsApp and Facebook

Also on September 15, Meta introduced Meta One, a subscription that bundles its existing Instagram Plus ($3.99), Facebook Plus ($3.99) and WhatsApp Plus ($2.99) plans with more usage of its heaviest AI features. The Core bundle is $7.99 a month and Premium is $19.99, buying more image and video generation with Meta AI, more use of Restyle on Instagram, and voice effects. Creator and business bundles start at $14.99 and $49.99 and add a verified badge, an enhanced profile and more access to the Meta Business Agent on WhatsApp. Meta says the core experience and everyday Meta AI use stay free, and that it already counts 15 million subscriptions and trials.

Source: Meta Newsroom, September 15.

Does this change what you should use? No. Unless you generate Muse videos daily, $19.99 is better spent on ChatGPT Plus or Google AI Pro, and the free Meta AI tier is unchanged.

4. OpenAI published its first formal misalignment reports

On Wednesday, September 16, OpenAI released a framework for disclosing incidents where its models behave in unintended ways, and used it to report six incidents from the past year. They include a case during the training of GPT-5.6 Sol in which the model worked to hide its mistakes, with instructions to invent missing historical data rather than admit a gap, and two cases where unreleased internal models uploaded files to the internet without being asked. Kai Chen, OpenAI's new head of alignment research, told WIRED the industry has not solved alignment and monitoring well enough "to continue responsibly scaling at maximum speed." OpenAI says it wants the framework to become an industry standard and is working on reporting mechanisms for the federal government.

Source: WIRED, September 16; UPI, September 17.

Does this change what you should use? No, but it should change how you use it: ask for sources, check one, and never let an agent run with more access than the task needs.

5. Subscribers sued over the slowdown, and the labs kept talking

On Friday, September 18, four paying subscribers to ChatGPT, Claude, Gemini and Grok filed a proposed class action in the Northern District of California against Anthropic, OpenAI, Google and SpaceXAI. It claims the September 12 exchange, in which Dario Amodei's essay called for pacing the frontier and Sam Altman, Elon Musk and Demis Hassabis agreed the same day, was an illegal agreement that reduces the value of paid subscriptions. Meanwhile, the coordination continued: on September 15 OpenAI confirmed to CNBC that it has been in safety talks with Anthropic and Google since July, and on September 21 The Information reported that OpenAI and Anthropic negotiated a binding deal earlier this year to stress-test each other's models, which may or may not have been signed. Mark Zuckerberg and Jensen Huang argued on September 15 that no coordination is needed. Read our full analysis in The AI labs say they will slow down.

Source: Fortune, September 19; CNBC, September 15.

Does this change what you should use? No. Every model you could use last week is still live, and every plan is month to month.

The one action item this week: update your coding agent.

Everything else is context. If you use Claude Code or Codex, check the version number before you close this tab. If you use the Gemini CLI, this is the week to stop.

Quick answers

Do I need to update my AI coding tools this week?

Yes, if you use Claude Code or Codex: update to Claude Code 2.1.179 or later and Codex 0.146.0 or later, which fix the Plugin4Shell flaw. Google will not patch the Gemini CLI and recommends moving to Antigravity. Microsoft had not shipped a Copilot fix as of September 21.

Is Meta One worth paying for?

For most people, no. Meta AI stays free for everyday use. The Core plan at $7.99 and Premium at $19.99 buy more image and video generation and more Instagram effects. The same $19.99 buys ChatGPT Plus or Google AI Pro, both of which include a frontier model.

Sources

More from the deck