The Signal — March 28, 2026
Anthropic confirms Claude Mythos after a 3,000-document leak. Apple prepares to open Siri to rival AI chatbots. Mistral releases open-weight voice cloning from three seconds of audio.
A misconfigured content management system, a 3,000-document data spill, and now Anthropic has confirmed the obvious: Claude Mythos is real, it's in testing, and the company calls it a "step change" in capability.
Anthropic's Next Model Leaks Through Its Own Website
Fortune reporter Bea Nolan discovered nearly 3,000 internal Anthropic documents sitting in an unprotected, publicly accessible data store on the company's website. Among them: draft blog posts describing a new model called Claude Mythos, described as "larger and more intelligent than our Opus models" with "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity."
When we flagged this story yesterday, it rested on a single source. Since then, Anthropic has confirmed to Fortune that it is "developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity" and considers it "the most capable we've built to date."
The leaked drafts reveal two competing name candidates. One version calls it Mythos, another calls it Capybara, but the Capybara version still contains "Claude Mythos" in the subtitle. Both drafts use identical justification for the name, saying it evokes "the deep connective tissue that links together knowledge and ideas." Anthropic told Fortune the documents were "early drafts of content that were being considered for publication."
What makes this more than a naming curiosity: the drafts say the model is "currently far ahead of any other AI model in cyber capabilities" but warn it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." Anthropic is planning a slower, phased rollout starting with early-access customers focused on cybersecurity applications, with API access expanding gradually. The drafts also acknowledge the model is "very expensive for us to serve" and say the company is working to make it efficient enough for general release.
The timing matters. OpenAI is reportedly preparing its own next-generation model, codenamed "Spud," and both companies are positioning for IPOs later this year. The leak gives Anthropic an unplanned preview that doubles as competitive signaling. Whether they wanted it or not.
Sources: Fortune (exclusive) · Fortune (follow-up) · The Decoder · Mashable · Quartz
Apple Plans to Open Siri to Rival AI Chatbots
Apple is preparing to let third-party AI services plug directly into Siri, according to a Bloomberg report published March 26. The change, expected in iOS 27, would allow users to route queries from Siri to services like Google's Gemini or Anthropic's Claude, moving beyond Apple's current exclusive arrangement with OpenAI's ChatGPT.
The mechanics: chatbot apps installed through the App Store would integrate with Siri and Apple Intelligence, letting users choose which AI handles each request. Bloomberg's sources say Apple could also take a cut of subscriptions sold through these third-party integrations.
This is Apple doing what Apple does best. Rather than building a frontier model (which it hasn't), it positions the iPhone as the default interface for whatever model you prefer. The approach carries an implicit admission that Apple Intelligence alone isn't competitive with dedicated AI labs, but it also means Apple doesn't need it to be. If every AI service needs to go through Siri to reach a billion iPhone users, Apple still controls the most valuable chokepoint in consumer AI.
Apple is expected to preview the features at WWDC in June. Plans could still change.
Mistral Drops Open-Weight Voice Cloning That Needs Three Seconds of Audio
Mistral released Voxtral TTS on March 26, the first open-weight text-to-speech model from a major AI lab. It clones voices from as little as three seconds of reference audio across nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic.
The model is available under open weights, meaning anyone can download, run, and modify it locally. That distinction matters for voice synthesis specifically. Commercial TTS services like ElevenLabs require API calls, usage tracking, and trust in the provider's safety guardrails. An open-weight model puts the capability and the responsibility in the same hands.
At 4 billion parameters, Voxtral is lightweight enough to run on a smartphone, according to VentureBeat. Mistral says it was trained on a large multilingual speech dataset and built for "global applications." The company claims it beats ElevenLabs on quality benchmarks, though independent testing hasn't confirmed that yet.
Voice is rapidly becoming the next modality where open-weight models challenge proprietary APIs. Text generation got there first. Image generation followed. Voice cloning from three seconds of audio, running locally, is the kind of capability that was science fiction two years ago and is now a download.
Sources: TechCrunch · SiliconANGLE · Forbes
On the Editor's Desk
We held the Claude Mythos story yesterday when Fortune was the only source. By this morning, Mashable, The Decoder, Quartz, and CoinDesk had all confirmed the details independently, and Anthropic went on the record. That's the threshold. Single-source leaks stay on the desk until they don't.
YC's Winter '26 Demo Day (March 24) put 190 startups through their paces. TechCrunch highlighted 16 standouts across law, transportation, and healthcare. We're not covering it today because the event itself is four days old and the roundups are summaries, not news. If any of those 16 companies ship something worth tracking, we'll be there.
David Sacks officially stepped down as the White House AI and Crypto Czar. Multiple outlets reported on it this week. We're tracking but not leading with it because the more interesting story is what happens to AI policy coordination next.
