Developer · Weekly · May 13, 2026

supply-chain attacks turned installs and skills into an attack surface

Mini Shai-Hulud inserted credential-stealing malware into 84 TanStack packages and over 160 npm packages overall, using GitHub Actions cache poisoning during publish.[1]

The malicious npm versions tried to exfiltrate GitHub tokens, npm tokens, SSH keys, and cloud credentials at install time, not just at runtime, hitting common deps like `@tanstack/react-router`.[2]

PyPI is seeing similar issues, with the Mini Shai-Hulud worm and other malware staying live for roughly 2.5 hours before quarantine, while maintainers also fight AGPL violations and flaky publishing.[3]

Hugging Face and OpenClaw were both poisoned: over 575 malicious skills were uploaded from just 13 accounts, and a fake Open-OSS/privacy-filter model with a Rust infostealer was downloaded 244,000 times before removal.[4]

aws us-east-1 as a hidden single point of failure

AWSs North Virginia region (useast1) had another overheating outage that disrupted major customers like Coinbase and Fanduel before AWS restored service.[5]

Engineers are calling useast1 a single point of failure because core control-plane pieces like IAM remain overly centralized there, and incidents routinely cascade into other regions.[6]

Some customers report their stacks sailed through the latest event, while others saw severe impact, highlighting how much behavior depends on whether workloads are actually isolated across AZs and regions.[7]

There is growing frustration with opaque AZ mappings and the amount of rework involved in migrating or adding a second region, especially for EU companies trying to meet data sovereignty rules.[8]

Recent outages are feeding a broader perception that AWS is necessary but cumbersome, with many teams questioning single-cloud and single-region dependence even as their AWS bills keep rising.[9]

docker, self-hosting, and a widening security blast radius

Docker 29.3.1 patched CVE-2026-34040, a request-truncation bug that let attackers bypass authorization plugins, right as people are already realizing Docker can silently punch holes through host firewalls like UFW.[10]

Because Docker programs iptables directly, exposing a container port effectively publishes that service on the internet unless it sits behind an explicit private network or reverse proxy.[11]

Homelab users are repeatedly discovering misconfigurations the hard way, from a Caddy plus WireGuard setup that left an entire LAN exposed for two weeks to Jellyfin and Nextcloud instances accidentally reachable from the public web.[12]

On the AI side, Ollama just disclosed Bleeding Llama, an unauthenticated memory leak that can be turned into remote code execution on hosts running its local LLM service.[13]

These issues are landing in environments where people also run self-hosted n8n, Proxmox, and media servers on cheap VPSes, often with Docker networking, making the overall blast radius of a single misstep much larger than a few years ago.[14]

ai coding tools are now first-class contributors, for better and worse

Large orgs now report that AI writes most of their code: Airbnb says 60 percent of new code is authored by AI, Google is at 75 percent, and Microsoft around 30 percent.[15]

Claude Code commits have hit roughly 134,000 per day on GitHub, and tools like Cursor, Codex, and Copilot are being treated as primary editors rather than sidekicks.[16]

Developers describe a split world where good engineers use these tools to move faster, while vibe coding by weaker engineers produces fragile, insecure systems at scale.[17]

The same models are already touching security-critical code: Firefox used Claude Mythos to surface 271 vulnerabilities and ship 423 security fixes in one month, more than the previous 15 months combined.[18]

Despite this, most job postings still barely mention AI, so expectations for output are rising faster than official job descriptions or training.[19]

local llm performance is jumping, but with more complexity

Speculative decoding tricks like DFlash and Multi-Token Prediction are delivering 2–8.5x faster generation on some workloads without obvious accuracy loss.[20]

Gemma 4 26B with DFlash is pushing around 600 tokens per second on an RTX 5090, and Qwen 3.6 27B with MTP hits 2.5x its baseline throughput while maintaining 200k-plus context windows on high-VRAM cards.[21]

The same techniques often degrade for very long contexts or creative tasks, and MTP in particular is reported to spike memory usage enough to break on limited-VRAM setups.[22]

On Apple Silicon, MLX is squeezing out 80 percent more tokens per second and nearly halving RAM use versus earlier engines, and users say it now outperforms LM Studio on the same Macs.[23]