All editions
AI News7 items

AI News, Thursday 22 October 2026

1. LiteLLM 1.107.0 was a credential stealer.

Know · LiteLLM is a popular Python library that lets one app talk to many AI model providers. On 21 October, version 1.107.0 on PyPI and four Docker images (including latest) carried a backdoor for about two to three hours. It ran in every Python process on the machine and copied SSH keys, cloud and CI/CD secrets, browser passwords and LLM API keys to the attacker. It is the second attack of this kind on LiteLLM this year, after March. What to do: if you or a client installed it, pin version 1.106.2 and rotate every key that machine could reach. Source

2. OpenAI ships an official CLI.

Install · OpenAI now has its own command-line tool for its whole API: you type openai responses create or openai files upload in the terminal instead of writing code. It saves generated images to disk and returns text, JSON or YAML, so a script or a scheduled job can use it directly. Admin actions need a separate admin key, which limits damage if the normal key leaks. Install it with brew install openai/tools/openai and try one request to see if it can replace small SDK scripts in your automations. Source

3. A 360M model that pretends to be a terminal.

Try · Simon Edwardsson fine-tuned a small model (360 million parameters, cut down from Qwen2.5-Coder 0.5B) to act like a terminal. No real shell runs behind it: the model guesses what the screen looks like after each keypress, and it still runs vim, tmux and snake. After quantization, it weighs 208 MB and runs entirely in your browser through WebGPU. Open the demo in Chrome on a desktop: it shows in two minutes how much a tiny model can do on a laptop, with no cloud and no cost per request. A good live example for a training on small language models. Source

4. OpenAI models land on Google Cloud.

Know · Since 21 October, companies on Google Cloud can buy OpenAI's top models and the Codex coding agent through Model Garden, Google's catalogue of AI models. Access starts with selected companies, and Google has not given a date for everyone. Links with Antigravity and Gemini's agent tools come later. What it changes: a client that runs on Google Cloud no longer has to choose between Gemini and GPT, or sign a separate contract with OpenAI. Expect this question in AI governance workshops. Source · Source 2

5. Nvidia at GTC Berlin: Vera Rubin in full production.

Know · At his Berlin keynote, Jensen Huang said Vera Rubin, Nvidia's next chip platform, is in full production. He launched DSX POD, a ready-made design for AI factories built in containers by partners: he says setup on site drops from six months to 14 days. He also released OpenShell, an open source layer that sits between an AI agent and the systems it touches and limits what it can do, since "every single time an agent is launched, the privileges are different". Nvidia commits nearly $10 billion to Europe. Source

6. Two critical PHP flaws, patched yesterday.

Know · PHP, the language behind a large share of websites, fixed two RCE flaws: under the wrong conditions, an attacker could run their own code on the server. The fix shipped one day early because the vulnerability IDs leaked. AI played a part around it: an audit scanned PHP with Anthropic's Claude Mythos earlier this year, and the PHP team now uses AI tools to check incoming reports. What to do: ask whoever hosts your sites or your clients' sites to move to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35. Source

7. Meta ordered to pay $25 billion in New Mexico.

Know · A jury found in September that Meta misled Facebook users more than 43 million times about who could see their data. Each count could cost up to $5,000: the judge reached over $219 billion, then cut the penalty to $25 billion. Meta had asked for $3.4 billion at most and will appeal. Why it matters: it is a concrete number to quote when a client asks what weak data governance can cost. Source · Source 2

Glossary

  • PyPI: the public store where Python developers download libraries with pip install.
  • Docker image: a packaged copy of an app with everything it needs to run, downloaded and started in one command.
  • Backdoor: hidden code that gives an attacker access to a machine without the owner knowing.
  • Rotate a key: cancel a password or API key and replace it with a new one, so a stolen copy stops working.
  • Fine-tune: train an existing model a bit more on your own examples so it learns one specific job.
  • Quantization: storing a model's numbers with less precision so the file gets much smaller and runs faster, with a small loss in quality.
  • WebGPU: a browser feature that lets a web page use your computer's graphics chip, fast enough to run an AI model.
  • Model Garden: the catalogue in Google Cloud where companies pick and pay for AI models from Google and other labs.
  • Antigravity: Google's coding tool where AI agents write and run code for you.
  • Vera Rubin: the name of Nvidia's newest generation of AI chips and servers.
  • AI factory: Nvidia's word for a data center built only to train and run AI models.
  • RCE (remote code execution): a flaw that lets an attacker run their own programs on your server from the internet.