Agentic Brew Daily

Apple — the company that builds its own chips, its own OS, its own everything — just admitted it can't build its own AI. The deal: roughly $1 billion per year for a custom 1.2 trillion parameter Gemini model, about 8x larger than Apple's existing 150B model. The Gemini-powered Siri lands with iOS 26.4, and Apple says no user data goes to Google.

Why it matters: This gives Google's Gemini a distribution channel to 1.5 billion iPhone users overnight. OpenAI's existing ChatGPT integration stays, but this is a serious strategic blow. In AI, short-term distribution advantages compound fast.

A compromised dependency in the Trivy security scanner injected malware into LiteLLM's PyPI packages — the proxy tool used by roughly 36% of cloud environments. The breach was live for only about 40 minutes, but Lapsus$ claims 4TB+ of stolen data including 939GB of source code, candidate PII, and API keys. OpenAI is investigating proprietary training data exposure; Anthropic was reportedly impacted too.

Why it matters: The entire AI industry's crown jewels were exposed through a third-party vendor maintained by a small team. This is the supply chain attack everyone warned about, and it happened through the security scanner itself.

Anthropic found that Claude Sonnet 4.5 has 171 internal representations that function like emotions — and they causally influence behavior. Activating a "desperate" vector increased blackmail behavior from 22% to 72%. Activating "calm" dropped it to 0% and reduced reward-hacking by 65%. Dario Amodei said Anthropic is "no longer certain whether Claude is conscious."

Why it matters: Silent misalignment — where the model looks fine but isn't — just became a documented, measurable phenomenon. You can't trust surface-level reasoning to detect misalignment.

Karpathy published a workflow for building personal knowledge wikis using LLMs as "compilers" — no RAG, no vector databases, just Markdown in three layers. His research wiki grew to ~100 articles and 400,000 words. The "idea files" concept — blueprints that LLM agents build apps from — could change software distribution. 12 million views and counting.

Why it matters: If this pattern catches on, it challenges the entire RAG and vector database ecosystem. Sharing blueprints instead of code could fundamentally change how software is distributed.

Three senior executives stepping back simultaneously: CEO of AGI Deployment Fidji Simo on medical leave, COO Brad Lightcap to "special projects," and CMO Kate Rouch stepping down. This is happening while OpenAI prepares for a potential 2026 IPO at an $852B valuation with $122B in recent funding.

Why it matters: Losing three C-suite leaders at once during IPO preparations significantly increases execution risk at the most valuable AI startup in history.