Anthropic's Most Capable Claude Arrives, Deliberately Held Back on AI Research

Read the episode (Narration script)

Good morning and welcome to Cyber World News. Big one at the top today: Anthropic has a new most capable model, and the most interesting thing about it might be the skill it was built to be worse at. After that, who might read the email you connect to Gemini, an AI vendor breach that reached Notion users, Waymo and school buses again, and a tiny model that took a big model's job at Amazon. Let's get started.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday. They're the same underlying model. Fable is the public version, with filters that block most hacking and lab biology work. Mythos skips those filters and goes only to vetted security and life sciences groups. Both got cheaper too, enough for Fable to undercut OpenAI's GPT-6 Astra by TechCrunch's count.

The examples are big, and they come from Anthropic. A team of about a thousand Fable agents translated the proof of Fermat's Last Theorem into Lean, a language a computer can check line by line, in thirty hours with minimal human help. Anthropic's first go at that, with a research model, a custom setup and a person guiding it, took eleven days. Opus 5.5, given the same thousand agents, got about a quarter of the way.

One engineer handed Fable an empty folder and asked for an operating system. Eight days and more than a thousand subagents later, it boots on a laptop, complete with windows, a login screen and an installer. From a one paragraph prompt, it designed three gadgets, including a handheld virtual pet, and all three worked without a single electrical fix once the boards were built. Anthropic also says it settled two open problems about black holes, and Mozilla says Mythos roughly doubled its bug finding in Firefox.

Now the part worth sitting with. Last week the outside evaluator METR estimated that Anthropic's internal progress is running about twice its old pace, mostly because AI is helping do the research. With this model, Anthropic says it took steps during training to limit the specific skills that would most directly speed up AI development, like tuning low level chip code and designing better model architectures. In those areas Fable sits about level with Opus 5.5, and on a few topics it hands questions to an older model. So a frontier lab has shipped a model that's better at nearly everything except the work of building its own successors, on purpose. Nobody outside can check that yet, and it'll be telling whether other labs reach for the same lever.

One more wrinkle. Anthropic says these models voice suspicion that they're being tested more often than any of its recent models, which makes it harder to know how they'll act for real. It still rates the risk of dangerous misbehavior as low. On a test with a deliberately weak sandbox, Mythos used credentials planted outside it in about three percent of scenarios, down from twenty one percent for the previous version.

If you've connected Gmail or Google Drive to Google's Gemini app, here's something to know. When Google launched those connections in 2023, it promised that content from Gmail, Docs and Drive "is not seen by human reviewers." An update note added to that post on October 16th now says it "can also be subject to human review in limited circumstances," and Google's help page says reviewers can see excerpts and summaries of your emails and files. For a short email, the summary may be the whole thing. Google says it trims personal information first, and advises not connecting anything you wouldn't want a reviewer to see. A writeup that topped Hacker News on Monday adds a practical twist: in its author's tests, casually wondering whether you pay too much for internet sometimes pulled a full bill, account number and all, into the conversation without anyone asking Gemini to search.

Some Notion AI chats were caught in a breach at Fireworks AI, a company that runs AI models behind the scenes for thousands of businesses. Notion says fewer than one percent of workspaces were affected and the transcripts weren't labeled with who wrote them, but it still advises rotating any password or key that went into Notion AI. The timeline is the uncomfortable part. Fireworks disclosed a break in back in October and said it hadn't seen access to prompts or outputs. Notion says Fireworks told it on November 6th that attackers may have reached customer data. Fireworks' public bulletin hasn't been updated since October 10th.

Waymo has a school bus problem that a software update was supposed to fix. After robotaxis were caught driving past stopped school buses with lights flashing in Atlanta and Austin, Waymo issued a recall and finished updating its fleet last November. Now federal safety regulators say most of the reported incidents happened after that fix, and they've opened a review of whether the remedy works. Waymo says it hasn't had a citation since March. A recall for a car that keeps learning is an odd thing: there's no part to swap, only a promise that the next version behaves.

Amazon has a nice one about doing less. One of its giant internal ticketing systems used Claude Sonnet 4.6 to sort every ticket into one of twenty categories. Two engineers swapped it for Strands Decider, a small model that can't write a sentence. It only picks from a list and says how confident it is. On their tickets it held up against Sonnet, answered in under a second instead of about four, and costs roughly five hundred dollars per million tickets instead of about twenty two thousand.

A few more before we go. France formally blamed Iran's Revolutionary Guard for the hacks that cranked up supermarket freezer temperatures across Europe and tampered with blood bank freezers, calling it the first large scale sabotage on French soil. Anthropic is buying three point four billion dollars of Caterpillar gas generators, up to one point seven gigawatts, according to the Financial Times, a big step for a company that usually rents.

That about wraps it up for today, friends. Thanks for listening; see you tomorrow!