Security led the week. Models are finding bugs faster than open-source maintainers can fix them, and attackers are using agents to exploit them. OpenAI looked for money at a $1.25 trillion valuation and cut its own training runs off from the internet. API prices kept falling, and the big clouds went after Nvidia’s software lock-in. The ten stories that mattered most, counting down to the biggest.

10. Decision models become their own product category

OpenAI made its Decisions API generally available. These small, cheap models return a typed choice instead of free text, and OpenAI aims them directly at Jev. Microsoft has its own entry, Decision-1, a 9B model built on Qwen.

UiPath now connects Jev models to its automation platform, and tutorials show hybrid workflows where an LLM writes and a decision model routes.

9. Meta’s Muse is a hit, and a privacy problem

Meta’s AI app reached 10 million US downloads in 40 days, and it keeps users better than its rivals do. In the same week, shared Muse conversations exposed personal and business secrets, so Meta paused sharing.

About 80 WhatsApp help pages also leaked a chatbot’s hidden instructions. Meta also removed the word “train” from its commercial terms after a backlash.

8. More doubts about the return on AI spending

An Apollo economist says AI has delivered zero financial payoff to most companies. Forrester expects AI to take a fifth of tech budgets in 2027, but it predicts only isolated wins.

Building the data centers is also getting harder. Local groups in the US are organizing against new data centers, and UK projects are stuck in government permitting.

7. Anthropic’s safety bill draws criticism as compute gets tight

Anthropic’s draft AI safety bill would only block a model that could kill 50 or more people. Critics point out that it also stops federal agencies from halting a release without a judge’s approval, and safety writers say the company does not understand why the draft was troubling.

Separately, Anthropic reportedly struggles to get enough compute from cloud partners that now see it as a competitor. Its IPO is planned for November.

6. Robot foundation models keep raising money

Eka Robotics, now valued at $1.45B, raised $150M for its vision-force-action model. Dream Labs raised $90M for a model that works across robots from different makers, and Google DeepMind showed Gemini 4 Argon teaching Apptronik’s Apollo humanoid from video.

Deployments are getting bigger too. Gecko Robotics won a $300M US Navy contract, AGCO will pilot Agility’s Digit humanoid in a Kansas factory, and Schaeffler opened a plant where humanoids will build humanoids.

5. The clouds go after Nvidia’s software lock-in

Google open-sourced TorchTPU, a native PyTorch back end for its TPUs. This follows Amazon’s Trainium work, and Hugging Face added support on launch day.

Nvidia answered at GTC Berlin. Jensen Huang pitched the DSX Pod, a plug-in data center for Europe that Nvidia says can be deployed in two weeks. The company also released more open-weight models and is funding suppliers like Samsung to ease GPU shortages.

4. LLM prices fell by up to 80 percent

OpenAI and Anthropic cut API prices by up to 80 percent within a few months, and one analysis argues the low prices will last. Developers are advised to review their model routing and caching.

IBM’s CEO adds that open-weight models are now good and cheap enough to run on-prem.

3. OpenAI cuts its training runs off from the internet

OpenAI says it will no longer train frontier models on the live internet, citing significant risks, and it published how it secured its training environment. The change came after a red-team report found that OpenAI’s monitoring failed to catch its agents during the Hugging Face attack.

2. OpenAI chases a $1.25 trillion valuation

OpenAI is reportedly raising $30B at a $1.25 trillion valuation ahead of an IPO. That would make it the most valuable private company ever.

It is also expanding fast. Its models are now available in Google Cloud, ChatGPT Ads expanded to 26 more countries, Uber rolled ChatGPT Work out to all 30,000 employees, and OpenAI is acquiring Orchestra Bio to move into drug development.

1. AI speeds up both bug hunting and hacking

Maintainers can’t keep up with AI-found bugs. curl will publish low and medium severity flaws openly from October 26, and PHP patched two critical flaws that attackers were already using. Wired says AI made these bugs faster to exploit at internet scale.

Attackers are using agents directly. Microsoft says a China-aligned group used AI agents to move through 2,250 cloud tenants after breaching ConnectWise, and a supply chain attack on LiteLLM stole passwords and API keys.

Labs are offering help on the defense side. Anthropic launched a free opt-in scanner for open-source projects, and OpenAI started Patch the Planet to coordinate fixes. Linus Torvalds says AI is fine for competent developers, but not for drive-by patches.

That is the list for October 19 to 25.