|
AI Builders Digest
Thursday, November 12, 2026
|
|
Security week continues. Yesterday Microsoft showed how a dormant-neuron backdoor could compromise any agent you deploy. Today, Vercel dropped numbers showing how many vulnerabilities are already sitting in your repos right now, waiting. The AI-finds-AI-problems era is officially open for business.
|
|
---
|
|
01
|
Vercel's security agent found 57,897 vulnerabilities in one week — and patched 14,000 of them automatically
|
|
|
Vercel CEO Guillermo Rauch posted the first-week numbers for Vercel Security Review: 57,897 vulnerabilities identified across codebases, 6,029 rated critical or high, and 14,665 already patched by the agent without human intervention. The tool works on any repository, regardless of where you deploy.
|
Why it matters: Pair this with yesterday's Microsoft Lazarus research and you get the full picture. The same AI agents being sold as productivity tools are also the fastest security auditors anyone has ever shipped. If your engineering team hasn't run something like this yet, the question is no longer whether you have critical vulnerabilities. It's how many.
|
|
Source →
|
|
---
|
|
02
|
Box and Anthropic give Claude full read/write access to enterprise files
|
|
|
Box CEO Aaron Levie announced a partnership with Anthropic that lets Claude read and write directly to Box content via MCP. Agents can now pull a contract, edit it, and save it back to Box without a human in the loop. Levie's framing: agents will interact with enterprise files 100 times more than people do today.
|
Why it matters: Most enterprise AI demos stop at "the agent found the document." This extends the workflow to "the agent updated the document and filed it." For any team managing contracts, strategy docs, or compliance paperwork in Box, that's the difference between AI as a search tool and AI as a coworker with edit access. The liability question of who reviews what the agent wrote back is now your legal team's problem.
|
|
Source →
|
|
---
|
|
03
|
Google ships a Decisions API built for speed, not conversation
|
|
|
Google launched the Gemini Decisions API, powered by a dedicated model called Gemini Decision, designed for fast, calibrated choices across text and images. This is not a chatbot. It's a model purpose-built for routing, classification, and real-time judgment calls inside automated pipelines.
|
Why it matters: Most production AI systems spend the majority of their compute budget on decisions that don't require a full frontier model. A cheap, fast, purpose-built decision layer means the expensive model only runs when it has to. If this performs as advertised, it changes how developers architect multi-step agents, and it gives Google a foothold in the infrastructure layer, not just the model layer.
|
|
Source →
|
|
---
|
|
04
|
Google DeepMind's AlphaGenome 2 predicts how DNA variations affect disease
|
|
|
DeepMind released AlphaGenome 2, a model that predicts how specific DNA sequences and their variations affect individual human cell types and contribute to disease. It builds on the original AlphaGenome, pushing further into the gap between having a genome sequence and understanding what any given mutation actually does in a specific tissue.
|
Why it matters: Most drug targets fail because we don't understand why a genetic variant causes disease in one cell type but not another. A model that can predict those effects computationally shortens the front end of drug discovery from years to weeks for the right class of questions.
|
|
Source →
|
|
---
|
|
05
|
JioStar's chief architect on what AI actually looks like at 500 million users
|
|
|
OpenAI published a conversation with JioStar's Vijay Seshadri on running AI at scale for one of the world's largest streaming platforms. The focus: evaluations, guardrails, and the unglamorous work of taking AI from prototype to production.
|
Why it matters: Most AI case studies come from U.S. enterprise software buyers. JioStar serves 500 million users across a range of devices, languages, and connectivity conditions. What breaks at that scale is a different set of problems than what breaks at 50,000 seats, and Seshadri's hard-won lessons on evals and guardrails are worth reading before your team ships something that has to work everywhere.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|