Securing AI Agents — MCP, Agentjacking & the New Attack Surface (2026)

In June 2026, researchers took over AI coding agents — Claude Code, Cursor, Codex — with no phishing, no malware, and a public credential developers paste into their own apps. It worked 85% of the time. The vendor's response? "Technically not defensible. We're not fixing it." Welcome to the new attack surface nobody secured.
What you'll hear:
• What MCP (Model Context Protocol) is — "USB-C for AI" — and why it became an unreviewed internet-facing doorway
• The hygiene problem — Knostic verified exposed MCP servers; 100% had no authentication; 8,000+ reported by early 2026
• Agentjacking — how a public Sentry DSN let attackers poison the logs an agent reads (85% success, 2,300+ orgs)
• The pattern — prompt injection (OWASP's #1 AI risk), indirect injection, and last year's Black Hat zero-click CRM exfiltration
• The readiness gap — 83% of orgs deploying agentic AI, only 29% ready to secure it
• Four moves to secure agents — least privilege, sandboxing + human-in-the-loop, MCP gateways, agent creds as non-human identities
Sources: Tenet Security (Agentjacking) · Knostic · Cisco State of AI Security 2026 · OWASP Top 10 for LLMs · Zenity Labs (Black Hat 2025).
— Andrés Sarmiento
#AIsecurity #MCP #agentjacking #promptinjection #cybersecurity #TechUpdates