MCP Server Security: What My Own Audit Found
I ran my own production MCP fleet through the real 2026 attack classes — server chaining, supply-chain risk, exposure, OAuth flaws. The honest verdict.
I ran my own production MCP fleet through the real 2026 attack classes — server chaining, supply-chain risk, exposure, OAuth flaws. The honest verdict.
A live build log: wiring AI-scored deal signals into a VC fund's existing CRM and Slack — no new sourcing platform, no rip-and-replace.
Extraction isn't the hard part with board decks — trusting the number after is. What broke, what worked, building this for real portfolios.
Spec-driven development assumed building was expensive. AI made that assumption wrong — here's what actually replaces it.
No platform, no message bus — the real owned-vs-rented architecture and actual June 2026 costs behind a production AI agent swarm.
Two AI vendors broke in the same week — an export-control blackout and a quota bug. Here's what building for real AI agent reliability actually means.
‘Loop engineering’ means letting a system prompt your AI agents instead of you. An honest verdict from an operator who runs production loops every day.
Opus 4.8 launched. ChatGPT absorbed MCP. Claude Code shipped parallel agents. The model isn't the product anymore — the operating layer is. What shipped, what flipped, what to read.
$10k / month. Whatever your fund needs, shipped that month.
