Daily Technology: March 29, 2026
Meta's AI that predicts brain activity, Anthropic's cybersecurity superweapon, and agents running businesses
Today's technology news bridges neuroscience and AI: Meta released a model that predicts how human brains respond to stimuli, Anthropic revealed a model designed for vulnerability discovery, and we're seeing what happens when AI starts understanding us rather than just mimicking us.
Meta's TRIBE v2: AI That Predicts Your Brain's Response
Meta's Fundamental AI Research team released TRIBE v2, a foundation model that doesn't just process language or images — it predicts how your brain responds to them. This is a significant shift from AI that mimics human output to AI that models human cognition.
- What it does: Predicts fMRI brain activity when people watch videos, listen to audio, or read text
- How it works: Combines LLaMA 3.2 (text), V-JEPA2 (video), and Wav2Vec-BERT (audio) into a unified brain encoder
- Accuracy: 2-3x better than previous models at zero-shot brain prediction
- Scale: Trained on 451.6 hours of fMRI data from 25 subjects
- Surprise: Its predictions of "average brain" responses are more accurate than individual human recordings
"TRIBE v2 naturally learns five well-known functional networks: primary auditory, language, motion, default mode, and visual. Even though it's a deep learning 'black box,' the internal representations organized themselves into recognizable brain regions." — Meta FAIR research team
The model successfully identified classic brain regions — the fusiform face area (for recognizing faces), Broca's area (for language), and the parahippocampal place area (for locations) — without being explicitly programmed to find them.
This is genuinely different from typical AI releases. Instead of asking "can this model beat humans at a task?", Meta is asking "can this model understand how humans think?" The in-silico neuroscience angle means researchers can run brain experiments on computers before spending thousands on fMRI scans. The fact that the model discovers brain regions on its own suggests it's learning something real about human cognition, not just pattern matching. This could accelerate neuroscience research dramatically.
Claude Mythos: AI That Finds Your Security Vulnerabilities
Anthropic's leaked Claude Mythos model isn't just better at coding — it's specifically designed to find cybersecurity vulnerabilities. The company is so concerned about misuse that it's releasing to security defenders first, giving them a "head start" before attackers can exploit the capability.
- Capability: "Far ahead of any other AI model" at finding security vulnerabilities
- Benchmark: Dramatically higher scores on cybersecurity tests than Claude Opus 4.6
- Risk assessment: "Presages a wave of models that can exploit vulnerabilities in ways that outpace defenders"
- Release strategy: Early access to organizations focused on defense, not general availability
- Context: Anthropic already found 500+ high-severity vulnerabilities with current Claude
Anthropic engineers previously used Claude to find previously unknown vulnerabilities in production codebases — including deducing a flaw from a developer comment in a change log. This new model appears to take that capability significantly further.
Anthropic is being careful, but the cat's out of the bag. If they can build this, so can others. The "defenders first" approach is smart positioning, but it's also an admission: AI-powered vulnerability discovery is now a reality. Security teams should be using these tools yesterday. The model will likely be expensive (it's a new tier above Opus), but for organizations serious about security, the ROI is obvious.
Alibaba's AI Agents: Your New Digital Coworkers
While Western companies debate whether AI agents are ready for production, Alibaba is deploying them at scale. The company is providing AI agents to millions of merchants on Taobao and Tmall — giving them autonomous "digital employees" for customer service, pricing, and promotions.
- What agents do: Answer customer questions, distribute vouchers, adjust prices in real time
- Availability: 24/7 autonomous operation — no human supervision needed
- Scale: Millions of merchants, China's largest e-commerce platforms
- Timeline: Launching end of March 2026
- Quote: "Human and digital employees collaborating" within 1-2 years
Alibaba frames this as a response to the "OpenClaw" capabilities — referring to the broader agent automation trend that's emerged in 2026.
This is what "agentic AI" looks like in practice. Not a chatbot you ask questions, but a worker that runs business operations. The Chinese market has been faster to adopt autonomous systems — partly regulatory, partly cultural. Western e-commerce platforms (Amazon, Shopify) will face pressure to match this. For small merchants, this is genuinely transformative: a customer service team you can't afford becomes a line item. The question isn't whether this spreads, but how fast.
What Claude Mythos Actually Reveals About AI Progress
The Anthropic leak reveals more than a new model — it shows how frontier AI companies are thinking about capability and risk together. The "Capybara" tier above Opus, the cybersecurity focus, and the cautious rollout strategy tell us where the industry is heading.
- Pricing shift: A new tier above Opus means AI is getting more expensive at the top end, not cheaper
- Specialization: Mythos/Capybara focuses on coding and security — not general improvement across all tasks
- Risk-first release: "Defenders first" is becoming standard practice for dangerous capabilities
- Competition: OpenAI's Spud and Google's Gemini 3 are racing to match
We're seeing the end of "one model does everything" thinking. The frontier is fragmenting into specialized models: code models, security models, reasoning models, creative models. Each tier is more expensive and more dangerous. The companies releasing them are simultaneously proud of the capabilities and scared of them. That tension — "we built something powerful" vs. "we built something dangerous" — will define AI development for the next several years.
What This Means for People Working With AI
For developers: Claude Mythos's cybersecurity capabilities mean vulnerability scanning is about to become much more automated. If you're not using AI security tools, your code is being tested by attackers who are.
For researchers: TRIBE v2's in-silico neuroscience could dramatically reduce the cost of brain research. Virtual experiments before real fMRI studies.
For business operators: Alibaba's agent deployment shows what "agentic AI" actually looks like in practice. It's not replacing humans — it's giving small businesses capabilities they couldn't afford before.