Cloudflare proved remote Spectre against Workers, isolated-vm's glue code broke, and most EKS clusters still use the auth method AWS told you to stop. Three reminders that the boundaries you trust are thinner than you think.
DevSecOps
Cloudflare found a hole in its Dynamic Process Isolation and demonstrated a remote Spectre attack against its Workers production environment, reliably leaking up to 12 bits per second at 99% accuracy. The fix is already shipped: an improved DyPrIs plus the V8 Sandbox with in-process isolation. The team says it found no signs of active exploitation in three years, but the published paper is a reminder that speculative execution remains a cross-tenant risk even under coarse-grained timers and noisy shared hardware.
A type confusion in isolated-vm's C++ glue code lets untrusted code running inside a V8 Isolate corrupt host memory. Endor Labs escalated from a single ivm.Reference to full control-flow hijacking, a genuine guest-to-host sandbox escape. Notably, the V8 boundary held; the value-marshaling layer failed. With embedded AI agents like n8n, Activepieces, and Mastra on 1M+ weekly downloads, the 6.2.0/7.0.1 patches deserve an urgent rollout.
Cloudflare is introducing task-based OAuth consent. Client owners can mark specific scopes as optional, so on the consent screen users can deselect the ones they're not comfortable granting instead of facing a binary approve-everything-or-deny choice. That's a real win for MCP servers and other agents that routinely request broad permissions users rarely need. Existing apps work unchanged; if no optional scopes are requested, the flow looks exactly as before.
Tencent's Zhuque Lab open-sourced AI-Infra-Guard (A.I.G), a self-serve red teaming platform that scans AI infrastructure, MCP servers, agent skills, and runs jailbreak evaluation against LLMs. The latest release ships multi-turn attacks (Many-Shot, PAIR, GOAT) plus bytecode-bypass and charset-smuggling defenses, and its vulnerability library now holds 2,000+ CVE rules. It runs via Docker or one-click install, with a web UI and OpenClaw skill, worth a look for anyone shipping agents or MCP toolchains into production.
Production readiness, reframed: not "all tests pass" but an evidence-backed decision about the exact artifact you operate. This repo consolidates 25,359 checkbox lines from 215 documents into 16 navigable lifecycle phases plus a 10-track release sign-off path. The pitch is unusually honest, no checklist can prove zero defects, so each control gets a recorded Pass/Fail/Blocked disposition and an owner. Free for teams that want a defensible, auditable go/no-go process without buying an enterprise tool.
AWS deprecated aws-auth, the hand-edited ConfigMap that maps IAM identities to EKS permissions, in favor of a newer API-driven method. Yet a 2025 Kubernetes Security Report shows 81% of clusters still run the old one, against AWS's own guidance. That hard-to-audit wiring is the security illusion Kubernetes creates, and it's why roughly two-thirds of teams report K8s-related delays in shipping.
Governance has been the orphan of Terraform: teams stitch together separate policy languages and disconnected workflows to keep sprawling hybrid estates in check. HashiCorp's answer is tfpolicy, an HCL-based policy-as-code framework now in public beta inside HCP Terraform, so platform teams write and enforce rules in the same language as their infrastructure. The pitch: replace bolt-on policy tooling with lifecycle-aware governance that lives where the config already does.
Wiz audited six neocloud object stores (Nebius, Crusoe, Vultr, Lambda, Cloudflare R2, DigitalOcean) against AWS S3 and found compatibility masks real gaps. None match AWS's Block Public Access or IAM fine-grained controls; access keys lack the structured formats that secret scanners such as GitHub rely on. One tester found `delete-bucket-policy` deleted the whole bucket. Portability muscle memory is a security trap.
GitHub's latest CodeQL release turns its security scanner on the CI/CD pipeline itself, sharpening Actions queries for cache poisoning, output clobbering and untrusted checkouts. The fixes go after false positives hard, with alerts now pointing straight at the workflow expression at fault. JavaScript and TypeScript modeling also gains deeper Vue, promises and Fastify coverage. Sharper alerts matter far more than the 497-query default suite, since triage burns out first.
npm v12 now blocks lifecycle scripts during install unless a developer explicitly approves them, killing off the easiest supply-chain malware vector. Checkmarx warns attackers will simply shift malicious payloads into runtime code, and that approval fatigue could gut the checkpoint. The door is locked, but the house is made of glass.