AI Engineering Signal #164
Claude Code's "Don't Ask Again" feature exposed credentials in over 1,100 public GitHub files
Signals
Claude Code's "Don't Ask Again" feature exposed credentials in over 1,100 public GitHub files
audit any repo where Claude Code ran with suppressed prompts before rotating secrets.
Web
Goldman Sachs pulls Claude Code from bankers and traders
enterprise AI coding access now gated on secret-handling audits, not just capability.
Web
Llama-modchip runs 284B DeepSeek on 64GB RAM, no GPU
CPU-only inference at this scale changes local deployment cost assumptions for large open-weight models.
GitHub
Linus Torvalds sees AI-prompted kernel patch flood peaking
low-signal AI patch volume is declining; maintainer review burden may stabilize.
Web
Apple researchers use Mac mini with local Qwen to control Unitree G1 humanoid
91% success across 360 real-world trials sets a concrete bar for edge-deployed embodied agents.
Web
Mirendil trains frontier-class model with one developer and $10M compute
pretraining cost floor is dropping; team size is no longer a reliable moat signal.
Web
The Take
The Claude Code credential leak and Goldman's pullback are the same signal: agentic tools that suppress user prompts and touch filesystems need secret-scanning gates before deployment, not after. Meanwhile, the cost floor for both inference and pretraining is collapsing faster than enterprise security posture is adapting.
Subscribe
Related Signals