AI Daily Brief — 26 March 2026

The biggest Thursday of Q1. Fortune breaks an exclusive: a CMS misconfiguration at Anthropic exposed roughly 3,000 unpublished assets — and one of them was a draft blog post announcing Claude Mythos, an unannounced model the company calls “the most capable we’ve built to date” and “a step change in performance.” The draft warns Mythos is “far ahead of any other AI model in cyber capabilities.” Cybersecurity stocks tank the next morning. Mistral‘s open-weights Voxtral TTS hits mainstream press. Sakana AI‘s “AI Scientist” is published in Nature. Meta FAIR ships TRIBE v2 and Google DeepMind releases an empirically validated AI-manipulation toolkit.

Top stories

• Anthropic Mythos leak. Cybersecurity researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) discover ~3,000 unpublished Anthropic documents — images, PDFs, audio and a draft blog post — sitting in an unauthenticated public data store due to a CMS misconfiguration. The draft names Claude Mythos (the “Capybara” tier), describes “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity,” and warns the model “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace defenders.” Anthropic confirms the model exists, attributes the leak to “human error” in an external CMS tool, and says early-access customers are already trialling it. via Fortune
• Mistral — Voxtral TTS. Mainstream press picks up Mistral’s open-weights TTS model (4B params, 9 languages, ~90 ms time-to-first-audio, $0.016 per 1k chars). Mistral claims Voxtral was preferred over ElevenLabs Flash v2.5 in 68.4% of blind human listening tests; the weights ship publicly so enterprises can self-host without sending audio to a third party. via TechCrunch
• Sakana AI — “The AI Scientist” in Nature. Sakana’s paper describing a fully automated end-to-end research system — ideation, code, experiments, write-up, peer-review — is published in Nature. AI Scientist-v2’s output became the first fully AI-generated paper to pass a rigorous blind human peer review at the ICLR 2025 ICBINB workshop (avg score 6.33). Collaboration with UBC, Vector Institute and Oxford. via Sakana AI
• Meta FAIR — TRIBE v2. A predictive foundation model that takes sensory stimuli (video frame, audio, text) and predicts fMRI brain responses, trained on more than 1,000 hours of fMRI across 720 subjects. FAIR reports several-fold accuracy gains over linear encoders and a 70× spatial-resolution improvement over prior SOTA neural decoding. Model, code, paper and interactive demo are released. via Meta AI
• Google DeepMind — “Protecting People from Harmful Manipulation.” The first empirically validated, real-world toolkit for measuring how LLMs alter human thought and behaviour. Nine human-participant studies with more than 10,000 subjects across UK, US and India; finance showed higher manipulation susceptibility than health. All study materials released publicly. via Google DeepMind
• ByteDance — Seedance 2.0 goes global. The near-Hollywood-quality text-to-video generator expands internationally via CapCut’s Dreamina platform — Africa, South America, Middle East and Southeast Asia first (US excluded at launch), access limited to some paid users. The model debuted in China the previous month. via TechXplore

Money & infra

Two notable infra moves cross the wire on or close to 26 March. Arbor Energy + GridMarket announce a 5 GW zero-emission baseload power partnership for US data centres, starting 2029, using Arbor’s HALCYON oxy-combustion sCO₂ turbines (25 MW modular units, 3D-printed). Pure DC + SEGRO‘s joint venture wins planning approval for a £1bn, 56 MW, 24,000 m² AI data centre at Park Royal, West London (PUE ~1.2, closed-loop liquid cooling). Sub-text: TSMC CoWoS advanced packaging remains sold out through 2026, the binding constraint on NVIDIA Blackwell/Rubin and AMD MI400-class supply. via Data Center Dynamics

OpenAI & Sora shutdown

Coverage of Tuesday’s Sora shutdown announcement peaks. OpenAI says the standalone Sora web app will close 26 April, the API on 24 September. Reporting puts Sora’s daily burn near $1M with worldwide users collapsing from roughly 1M to under 500K. Altman reportedly “felt terrible” telling Disney CEO Josh D’Amaro, who signed a ~$1B character-licensing deal three months earlier. OpenAI’s framing: refocus on robotics and “automated researchers.” via Slate

Markets & policy

The Mythos leak sends cybersecurity stocks tumbling at Friday’s open: CrowdStrike -7%, Palo Alto Networks -6%, Zscaler -4.5%, with Okta, SentinelOne and Fortinet each around -3%. Anthropic is reported to be privately warning senior US officials that Mythos materially raises the likelihood of large-scale cyberattacks in 2026; early access is restricted to organisations focused on cyber defence while efficiency-tuning continues. Anti-AI protesters gather outside Anthropic’s San Francisco HQ the same day. The Bartz v. Anthropic $1.5B copyright settlement claim deadline sits four days away (30 March). via Yahoo Finance

Quiet corners

xAI’s Nikita Bier teases that Grok will be wired into X’s core recommendation algorithm — what he calls “the most important change ever made to X” — for rollout the following week. Microsoft AI is silent on first-party shipping; Thinking Machines Lab remains in stealth build-out off its 10 March 1 GW Vera Rubin commitment with NVIDIA.

By the numbers

• Anthropic leak surface area: ~3,000 unpublished documents exposed via CMS misconfig
• Cyber stocks day-after: CrowdStrike -7% · Palo Alto -6% · Zscaler -4.5%
• Voxtral blind-test win-rate vs ElevenLabs Flash v2.5: 68.4%
• TRIBE v2 fMRI training set: 1,000+ hours · 720 subjects · 70× spatial-resolution improvement over prior SOTA
• DeepMind manipulation toolkit study scale: 9 studies · 10,000+ participants across UK, US, India
• Sora shutdown economics: ~$1M/day burn · users ~1M → ~500K
• Pure DC + SEGRO London AI data centre: £1bn · 56 MW · 24,000 m² · PUE ~1.2

Compiled by AI Feed’s editor from publicly reported announcements on 26 March 2026.