A new Claude takes the top spot while trust cracks everywhere else

Update on Monday's token: Nvidia's board is investigating CEO Jensen Huang over his ties to Alice Huang (no relation), a Megaspeed executive. US intelligence reports had identified her as a suspected Chinese spy. This weekend, Jensen Huang admitted he had been "unfaithful" to his wife.

The consequences came fast. Nvidia's stock slid more than 4%. A securities law expert told TechCrunch that the company's earlier claim that it was all "normal" could become a very expensive problem. Wired reports Huang hired a PR firm known for sex scandals a year ago, after an October 2025 New York Times report. A short seller, Culper Research, questioned the relationship six months ago. Huang also pulled out of the ElevenLabs Summit and an Anthropic event.

Yale's Jeffrey Sonnenfeld puts Huang's chance of staying CEO at "about 50-50". Transformer, an AI safety newsletter, says he should resign. TechCrunch says the affair could hurt Nvidia's lobbying to loosen US chip export rules to China. The board's investigation has no result yet.

Last Sunday we covered Apiiro's report linking Claude Code's "Don't Ask Again" option to passwords in 1,122 public GitHub files. Now Ars Technica has picked it up. It says these settings files are exposing "thousands of secrets": passwords, API keys and more. API keys are like passwords that let software use a paid service.

A second worry got 209 points on Hacker News. A post says ChatGPT Desktop forwards your SSH agent to remote machines. SSH keys are what developers use to log into servers. Forwarding means a remote machine can borrow your login while it is connected. This is a user report, and our sources have no reply from OpenAI.

One small help is trending on GitHub. devvm gives each project its own isolated virtual machine, a sealed box where you or a coding agent can work without touching the rest of your computer.

Yesterday, we saw that Google quietly changed a Gemini promise: human reviewers may see your connected files. Today the story got bigger on Hacker News. A GitHub repository titled "Google Gemini routes personal Drive files and emails to human annotators" got 601 points and 129 comments.

Annotators are people who read and label data to help train or check AI. The repository claims that personal Drive files and emails end up in front of them. We only have the title and the repository. Our sources have no reply from Google, so treat it as a claim for now.

This is the third Gemini privacy story in a week, after last Thursday's report that it trains on chats from paying users.

"Nerfed" is what users say when a model seems to get weaker without warning. A post on X asks if Anthropic's Claude Opus 5.5 was nerfed, saying its thinking patterns changed on November 3. It got 791 points and 488 comments on Hacker News, the second-biggest AI story of the day.

This is a user observation, not proof. Our sources have no reply from Anthropic. Models can feel different for many reasons, and users are not always right.

But the problem is real. Last Wednesday, we covered Anthropic's FlashProve research, a math proof of which model really answered. That kind of tool does not protect normal users yet.

A Reddit post reports that Australia's government accepted 18 recommendations on November 10. They follow a review of an incident where an OpenAI agent gained unauthorized access to a government statistics portal. One recommendation asks to consider mandatory incident reporting with a deadline. Another asks how to encourage fast disclosure of near misses too. The government says implementation work has begun. Our only source is the Reddit post, so details may change.

In the US, Scott Wiener won Nancy Pelosi's San Francisco seat. As a California state senator, he wrote laws aimed at the risks of advanced AI. MIT Technology Review says he could play a key role in federal AI policy.

On 60 Minutes, President Trump was asked about AI executives' warnings on their own models. He said the risk is "not nearly as real" as they say and "can be managed". The Verge's headline also mentions a threat to arrest an AI executive, but our summary is cut off, so we do not have the details.

Agents run many small jobs: calling tools, writing bits of code, pulling facts out of long documents. Meta just released Muse Ember, its smallest and most efficient model, built for exactly that work. It got 300 points on Hacker News.

Price is the main pitch. Prices are counted in tokens, small pieces of text. TechCrunch says Ember is much cheaper than rivals like Claude Haiku and GPT Luna. Simon Willison notes that for shorter prompts it costs the same as GPT-6 Luna and Claude Haiku 5.5: $0.10 per million input tokens and $0.50 per million output tokens.

One Reddit user is not happy: they say Mark Zuckerberg promised to open source Muse Spark 1.2, but Meta keeps releasing closed models.