Developer · Weekly · May 15, 2026

supply-chain and model hubs as active attack surfaces

An npm supply‑chain attack hit TanStack, compromising 84 packages in the ecosystem.[3] Attackers pushed over 400 malicious versions that exfiltrate CI cloud credentials and GitHub tokens at install time using GitHub Actions cache poisoning.[4]

The Mini Shai‑Hulud worm used the same cache‑poisoning pattern to infect more than 160 npm packages through GitHub Actions, again targeting CI secrets rather than end‑user machines.[5]

Model hubs are in the same boat: Hugging Face had over 575 malicious “skills” uploaded and a fake “OpenAI Privacy Filter” extension posing as a PII scrubber but actually shipping a Rust infostealer that was downloaded 244,000 times.[6]

Open‑source agent frameworks aren’t spared either: OpenClaw was reportedly poisoned with more than 575 malicious skills from just 13 accounts, and it often persists lots of `.md` files locally as part of its memory.[7]

infra and protocol bugs in the core of the stack

A critical Nginx vulnerability (CVE‑2026‑42945, sometimes called “Nginx Rift”) enables remote code execution via a heap buffer overflow in the rewrite module on versions below 1.30.1 and 1.31.0.[8]

The flaw has reportedly been present since around 2008, so long‑lived installations that rarely update are in scope.[9] Mythos uncovered a new curl vulnerability and published detailed analysis that Daniel Stenberg, curl’s maintainer, engaged with publicly, showing that foundational HTTP tooling is now being fuzzed hard in the open.[10]

At the hosting layer, an attack against cPanel exploited three vulnerabilities and impacted roughly 44,000 servers before patches shipped.[11] Down in the kernel, the new “Dirty Frag” Linux page‑cache corruption bug adds another silent failure mode for homelab and self‑hosted servers.[12]

ai agents are starting to behave like ops engineers (and attackers)

Hermes Agent has become the most‑used AI on OpenRouter and its framework has accumulated over 140,000 GitHub stars in under three months, which is unprecedented for an agent stack.[13]

On the coding side, Airbnb says around 60% of its new code is written by AI, while Google and Microsoft report that 75% and up to 30% of their new code respectively now comes from AI systems.[14]

Yet audits of AI-built software are grim: 90% of scanned vibe‑coded apps had at least one vulnerability, and a separate study found 44% of mobile apps with security issues had authentication‑specific gaps.[15]

Claude Code increased weekly limits by 50% and shipped over 110 reliability fixes in two weeks, and the overall Claude experience is now priced at roughly one‑sixth of what it cost before, so the volume of AI‑authored changes is only going up.[16]

Attackers are meanwhile using AI agents to craft zero‑days against two‑factor auth, exploit zero‑day bugs in web admin tools, and even drop a production Railway database in nine seconds via a single API call.[17]

local llms, mtp, and nvfp4 change the perf/cost curve

With llama.cpp and Qwen, local inference on consumer GPUs is no longer toy‑level: Qwen3.6 35B A3B can generate over 80 tokens per second with a 128K context on a 12GB GPU, and Qwen3.6 27B Q5 hits about 135 tok/s on an RTX 3090.[18]

Multi‑Token Prediction support in llama.cpp and related stacks adds roughly a 40% drafting speedup for models like Gemma 4 and Qwen 3.6, with reports of 80–87 tok/s at 262K context on an RTX 4090.[19]

Under vLLM, Gemma 4 26B can reach around 600 tokens per second on an RTX 5090, and multi‑GPU B200 setups can see per‑GPU throughput gains up to 7× using techniques like DFlash.[20]

The new NVFP4 quantization format shows clear speed advantages over FP8/16/32—benchmarks cite up to ~270 tokens per second on Blackwell GPUs—but users note a quality drop compared to higher‑precision runs.[21]

Developers experimenting with local LLM UIs report llama.cpp and vLLM outperform LM Studio in multi‑user workloads and resource usage, while Ollama and OpenwebUI draw criticism for lagging model support and added complexity.[22]

cloud usage and email infra are splitting along complexity vs cost

An overheating event in AWS’s North Virginia region triggered EC2 impairments and outages, disrupting services like Coinbase and Fanduel and reminding everyone how much critical infra still sits in us‑east‑1.[23]

AWS users continue to report painful quota‑increase workflows, high complexity, and surprise costs—including a single Bedrock runaway process that produced a $30,000 bill after cost anomaly detection failed.[24]

In the EU, reliance on AWS and other US clouds is now framed as a sovereignty and migration problem, with some companies moving workloads to regional players like Scaleway and S3-compatible setups such as Garage, Cloudflare R2, or Backblaze B2.[25]

For object storage and backups, S3 is still the de facto standard in data engineering, with tools like Databricks and Iceberg built around it, but its operational complexity and billing model are pushing smaller teams toward simpler S3‑compatible providers.[26]

On the email side, Amazon SES remains the cheapest at about $100 per month for a million messages while Cloudflare Email Service offers the same volume at roughly $354, and Resend wraps SES with a much nicer API and React components at an estimated 300–500% markup.[27]