Daily News: March 29, 2026
Anthropic's Claude Mythos leaks, Alibaba deploys AI digital workforce, OpenAI preps Spud
Today's AI news brings major revelations: Anthropic accidentally revealed its most powerful model yet, Alibaba is deploying AI agents to millions of merchants, and OpenAI confirmed it's finished pretraining its next-generation model.
Anthropic's "Capybara" Leaked — A "Step Change" in AI
A configuration error in Anthropic's content management system revealed the existence of Claude Mythos, internally code-named "Capybara" — a new model tier the company describes as "by far the most powerful AI model we've ever developed."
- What leaked: 3,000 assets including a draft blog post announcing Claude Mythos/Capybara
- What it does: Dramatically better at coding, academic reasoning, and cybersecurity than Claude Opus 4.6
- Cybersecurity focus: "Far ahead of any other AI model" at finding vulnerabilities
- Release strategy: Early access to security defenders first, not general release
- Why it matters: Anthropic says it "presages a wave of models that can exploit vulnerabilities in ways that far outpace defenders"
"We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we're being deliberate about how we release it." — Anthropic spokesperson
The leak also revealed plans for an invite-only CEO summit in the UK, where Anthropic CEO Dario Amodei will showcase unreleased Claude capabilities to European business leaders.
This is Anthropic's response to OpenAI's GPT-5.4 and Google's Gemini 3. The cybersecurity angle is clever positioning — release to defenders first, claim moral responsibility, but the message is clear: this model is a weapon. The "step change" language echoes exactly what Morgan Stanley warned about last week. The Q2 2026 AI leap is already here.
Alibaba Deploys AI "Digital Workforce" to Millions of Merchants
Alibaba is launching a new service that provides AI agents to millions of merchants on Taobao and Tmall, China's largest online marketplaces. The move capitalizes on the "OpenClaw" frenzy to create autonomous "digital employees."
- What it does: 24/7 autonomous agents for customer service, voucher distribution, real-time pricing
- Timeline: Launching end of March 2026
- Scale: Millions of merchants across Taobao and Tmall
- The quote: "In the next one to two years, we expect the standard operating model to evolve into a collaboration between human and digital employees"
While Western tech executives debate whether AI agents are ready, Alibaba is deploying them at massive scale. This is what the "agentic" future looks like in practice — not a chatbot, but a digital worker running your business operations. The competitive pressure on Western e-commerce platforms just intensified.
OpenAI Finishes Pretraining "Spud" — Release Expected Q2
OpenAI has completed pretraining on its next major model, internally codenamed "Spud." CEO Sam Altman said development is moving "faster than most people expected." The typical post-pretraining timeline suggests a public release in Q2 2026.
- Timeline: Pretraining complete, Q2 2026 release expected
- Context: OpenAI has been "on defense" since Gemini 3.1 topped SWE-bench
- What's next: Fine-tuning, safety alignment, and optimization phase
- Rumor: Natively multimodal, potentially competitive with Claude Mythos
The AI arms race is now operating on quarter-by-quarter cycles. OpenAI, Anthropic, and Google are all releasing within weeks of each other. Spud vs. Mythos vs. Gemini 3 will define the Q2 benchmark wars. The real question: will any of these models meaningfully change the "AI is plateauing" narrative, or are we seeing diminishing returns at frontier scale?
Meta's TRIBE v2 Predicts Human Brain Activity
Meta's FAIR team released TRIBE v2, a foundation model that predicts how the human brain responds to video, audio, and text stimuli. The model achieves 2-3x better zero-shot prediction than previous approaches and opens new possibilities for "in-silico neuroscience."
- What it does: Predicts fMRI responses across video, audio, and text stimuli
- Architecture: Uses LLaMA 3.2 (text), V-JEPA2 (video), Wav2Vec-BERT (audio)
- Training: 451.6 hours of fMRI data from 25 subjects
- Key finding: Zero-shot predictions more accurate than individual human subject recordings
- Implication: Could run virtual neuroscience experiments before costly fMRI studies
This is a different kind of AI progress — not about replacing humans but understanding them. TRIBE v2 successfully identifies classic brain regions (fusiform face area, Broca's area) purely from training data, suggesting AI models can learn neuroscience without being explicitly programmed. The "in-silico experimentation" angle could accelerate brain research dramatically.
What This Means for Readers
Anthropic vs. OpenAI: Both companies are racing to release their most powerful models in Q2. If you rely on AI for coding or security work, expect significantly better tools within weeks.
The agent economy is here: Alibaba's deployment shows autonomous AI workers operating at e-commerce scale. Western platforms will face pressure to match.
Cybersecurity transformation: Claude Mythos's security focus signals that AI-powered vulnerability discovery is now a competitive advantage. Defenders need these tools as much as attackers do.