[ security notes ]
Practical research and launch notes on MCP security, prompt injection, build guardrails, and runtime protection.
2026 is the year the web started closing to AI crawlers. But robots.txt is a polite request, not an enforcement mechanism — and blocking a scraper just tells it to rotate infrastructure. Poisoning is the other lever.
Rule engines block what they recognize. ekmire's semantic analyzer runs second, on payloads that already passed the rules, and scores them 0.0–1.0 for prompt injection risk.
Build Guard's regex and AST engines catch known patterns in milliseconds. ekmire dev-audit catches the auth bypass and the missing tenant check that don't have a pattern at all.
ekmire's free tier has two modes. One routes AI analysis through ekmire's Groq endpoint. The other sends your payloads directly to your own LLM provider — bypassing ekmire's infrastructure entirely.
Every WAF you've used ran on stale rules. ekmire's threat feed turns NVD CVEs into live detection rules within hours and hot-reloads every running proxy in 60 seconds — without a restart.
Most teams add CI scanning after the first incident, and only run it on PRs. A daily-scheduled scan with ekmire and SARIF retroactively catches secrets committed before the detection rule existed.
MCP turns AI assistants into tool-using agents. That also turns every MCP server, manifest, tool description, and localhost bridge into a security boundary.
AI coding editors read comments, issues, docs, diffs, terminal output, and MCP responses. Attackers can write into those surfaces too.
Most small teams need runtime protection before they need a procurement cycle. A reverse-proxy sidecar gives them a practical path.
