AI Engineering Weekly #5

This is a production risk signal: if your CI/CD pipelines or tooling depend on GitHub repos that got caught in the blast radius, they may have silently broken. Anthropic says the overbroad takedowns were accidental, but the leaked code is already circulating.

TechCrunch

Anthropic's Claude Code source leak triggered a mass GitHub DMCA takedown

and they accidentally nuked thousands of unrelated repos in the process.

Web

attn-rot (TurboQuant-style activation rotation for better quantization) merged into llama.cpp

local inference just got meaningfully tighter; Qwen 2.5-27B reportedly fits on a 16GB GPU at near-Q4_0 quality with this technique, worth testing before your next hardware order.

GitHub

LiteLLM supply-chain attack hit Mercor and reportedly thousands of other orgs

if LiteLLM is in your stack as a routing layer, audit your dependency versions and review what credentials it touches.

Web

Intuit reports AI agents hitting 85% repeat usage by keeping humans in the loop

the production evidence here cuts against the "full autonomy" narrative; human-in-the-loop isn't a crutch, it's what drives retention.

Web

OpenAI quietly backed a lobbying group pushing age-verification requirements for AI

disclosed only after investigation; worth tracking as regulatory capture attempts around AI access controls accelerate.

Web

Iran war disrupting helium supply critical for chip fabrication and MRI-cooled quantum hardware

supply chain pressure on semiconductor fabs is real and not priced into most AI infrastructure roadmaps.

Web