Grafana 13.0.2: path traversal in Loki and Tempo, 7 fixes
Grafana 13.0.2 shipped June 2 with seven security-only fixes. Path traversal in Loki and Tempo are the most critical -- upgrade if you run 13.0.0 or 13.0.1.
AIOps & Observability
Articles in AIOps & Observability
Tiny-vLLM: LLM Inference in C++ and CUDA
Tiny-vLLM rebuilds vLLM's core inference algorithms in pure C++ and CUDA — no Python required. Here is what self-hosters and inference engineers can learn from reading 3,000 lines of clean, annotated code.
OTel for AI agents: same tool, new conventions
OpenTelemetry graduated at CNCF on 21 May 2026. It now has developing semantic conventions for LLM spans and agent spans. If you already instrument microservices with OTel, here is what changes when you add AI to the stack.
OpenTelemetry graduates in CNCF -- the debate is over
OpenTelemetry graduated in CNCF on May 21 — 12,000+ contributors, second-highest velocity. The debate about standardizing on OTel is now settled.
Forge: guardrails that make small models reliable
Forge adds guardrails to local LLM tool-calling in Python. Lifts an 8B model from ~32% to 84% on its eval suite — no model swap, just a reliability wrapper.
Needle: 26M model handles function calls, runs on CPU
Needle is a 26M parameter model built for function calling. Beats 270M+ rivals on benchmarks, runs on CPU without a GPU. Route tool dispatch locally and skip the API call.
seo-mcp: GSC and Bing data for your AI assistant
A small MCP server connecting an AI assistant to live Google Search Console and Bing Webmaster Tools data. Why I built it instead of using what existed.
Grafana 13 upgrade guide for self-hosted operators
Grafana 13 for self-hosted: Git Sync is GA, React 19 breaks community plugins, Unified Storage auto-migrates. Here is the list before you change the image tag.
Using AI Critique to Gate Agent Task Handoffs
I put a critique gate before every agent handoff. It strips vague assumptions before they reach downstream workers and produce wrong-but-confident output.
AI SEO audit tools: pass/fail yes, counting no
Cogny said 44. We fixed what it flagged, ran it again: 85. Then Lighthouse and SEOptimer in the same 30 minutes. Here is what each tool gets right and wrong.
Traefik 3.6.14: five CVEs, one auth bypass
Traefik 3.6.14 patches five CVEs. The critical one: CVE-2026-40912 bypasses ForwardAuth, BasicAuth, and DigestAuth via percent-encoded paths. Upgrade now.
Wazuh 4.14.5: the CVEs they hid in release notes
Wazuh 4.14.5 hid five security advisories in routine release notes. One is a pre-auth stack overflow on port 1514. Upgrade today — a PoC lands in late July.
AI agent telemetry: cutting prompt costs with data
A Python parser, Prometheus, and Grafana running locally gives per-agent cache hit rates and cost-per-session data. Here is what we did with it.
Wazuh 5.0: new agent protocol, new indexer
Wazuh 5.0 beta rewrites agent state handling and replaces Filebeat with a native indexer connector. 4.x agents need upgrading before they reconnect. What changes, what breaks, and when stable lands.