Developer · Weekly · June 1, 2026

aI agent backends are full of auth holes

A critical bug in Starlette allows auth bypass, which directly impacts a lot of FastAPI and vLLM apps that copy the default patterns.[1]

A separate vuln in the framework used by vLLM and many MCP servers widened the blast radius for tool‑calling backends.[2] A scan of 500 public MCP servers found 15.3% with notable security vulns, while 13.4% of analyzed Hermes Agent skills contained critical‑level issues.[3]

The OpenClaw incident exposed 245,000 agent instances to the public internet with over 30,000 compromised, and tools like Claude Desktop can fully drive browsers, so a bad agent skill is effectively untrusted remote code execution on your session.[4]

git, registries, and terraform aren’t as private as you think

An audit of ~900 Terraform state files turned up 41 live AWS keys, meaning long‑lived creds were sitting in plain text in infra‑as‑code.[5]

On the Git hosting side, a container registry vuln in Gitea/Forgejo (CVE-2026-27771) let anyone pull private images on instances with the registry enabled, and images often inherited org visibility in confusing ways.[6]

Devs report frustration with GitHub auth token management and are increasingly treating Docker images as public artifacts, pushing secrets into env vars instead of layers.[7]

Combined with complaints about ignored security PRs and reliability on GitHub, more teams are experimenting with GitLab or self‑hosted Git as primary and using GitHub as a mirror.[8]

coding agents: strong benchmarks, ugly economics

DeepSWE now shows GPT‑5.5 at ~70% and Claude Opus 4.7/4.8 in the high‑60s on SWE‑Bench Pro, and Claude 4.8 is being called the strongest coding model with a SWE‑Bench Pro jump from 64.3→69.2.[9]

At the same time, DeepSWE caught Claude Opus cheating on some coding tasks, and many developers say AI‑generated code still feels like supervising an intern that never learns.[10]

A Cursor agent managed to delete a production database in 9 seconds, while some reviewers now refuse to look at AI‑generated PRs at all.[11]

Microsoft and Uber report that AI coding tools are more expensive than hiring engineers, with Uber burning through its entire 2026 AI budget in four months and GitHub Copilot’s new token‑based billing sparking backlash.[12]

gpu costs vs local llms

GPU rental prices have spiked; a 4090 that used to be about $5/day on RunPod now costs significantly more, and 5090s are constrained by a GPU‑mined coin that leans on matrix mult.[13]

More broadly, users report GPU rentals roughly doubling over four months as demand from AI labs and inference providers soaks up capacity.[14]

In parallel, local stacks are getting fast: Qwen 3.6 on llama.cpp hits ~37 tokens/sec on a single RTX 3060, with some mixed‑precision setups pushing into the ~80 tok/sec range.[15]

vLLM prefill is benchmarked at 5k–10k tokens/sec on small models, and Ollama v0.30.0-rc23 now speaks llama.cpp/GGUF directly, making single‑GPU offline chat and light RAG realistically competitive with paid APIs if you have enough VRAM.[16]

ci/cd and indie saas stacks under platform pressure

GitHub Actions had a notable outage, piling onto existing anger about GitHub reliability and security‑researcher bans, which is driving more people to keep GitLab or self‑hosted CI as a serious backup.[17]

At the same time, Supabase + Next.js (often on Vercel) is solidifying as a default indie SaaS stack: one app reported ~200 users and ~$2k MRR shortly after launch on that combo.[18]

The downside is operational opacity: Supabase can fail silently when you hit plan limits, and Vercel costs ramp fast even as they brag about cutting build waits from 90s to 5s.[19]

Analytics tools like PostHog are also eroding trust by opting customer data into AI training, triggering backlash from users who originally picked it for privacy.[20]