Fetching from the wire…
Public story · 2026-07-08 · high
The same audit found 43% have command-injection holes and 79% store credentials in plain text.
Why now: Codersera pitches this checklist as the 2026 security baseline for MCP servers, not a nice-to-have add-on.
A security audit found that 40% of MCP servers run with no authentication at all, per Codersera. Anyone running one for internal tools or a customer-facing agent has three ways in. No login screen, an injectable tool input, or a credentials file an attacker doesn't even need to crack.
The audit lays out five fixes. Use OAuth 2.1 with mandatory PKCE, not optional. Validate the token's audience so one scoped for a single service can't get replayed against another.
Allow-list every tool input instead of trusting what comes in. Block egress to private IP ranges so a compromised tool can't pivot into your internal network over SSRF. And never forward the client's token upstream to another service.
That's the rule Codersera says people miss, and it's the one that matters most. Passing a user's token to a downstream MCP server looks like normal auth delegation. It also means one compromised server inherits the permissions of every service it talks to. If you've chained MCP servers together, audit that chain before anything else on this list.
Most teams will fix the obvious stuff, the missing auth and the plaintext secrets, and skip the token-forwarding review because no scanner flags it. That's the hole that gets used first.
Each link below shares sources, entities, or timing with this story.
Censys counted 12,520 internet-accessible MCP services in June, most unauthenticated. Installing an MCP server is installing unvetted code with tool access, full stop. The concrete checklist before you trust one: require OAuth 2.1 + PKCE with token-audience validation, allow-l...
AWS's August 21 post stages agent tool governance as Connect, Control, Catalog and Harden, from one SSO-backed MCP endpoint for a 1-20 user pilot through identity-aware authorization with PII redaction and self-service tool publishing at 100+ users. It supports Cognito-backed...
The agent skills threat isn't isolated. The infrastructure layer is equally compromised. The Cloud Security Alliance's March 13 State of Cloud and AI Security report analyzed over 7,000 MCP servers and found 36.7% potentially vulnerable to server-side request forgery (SSRF). I...
BlueRock scanned over 7,000 MCP servers against 22-plus security rules. 36.7% carry potential server-side request forgery exposure from unrestricted outbound fetch, and 42% handle credentials insecurely. Their worked example is Microsoft's 85K-star Markitdown MCP server and it...
The new authorization spec, co-developed by Anthropic, Arcade, Microsoft, and Okta/Auth0, classifies MCP servers as OAuth 2.1 resource servers and standardizes audience binding via Resource Indicators (RFC 8707) and token exchange via RFC 8693. Aembit's writeup explains the ch...
upstash/context7 (60,590 stars) shipped @upstash/[redacted] on August 7 on the 2026-07-28 protocol revision. HTTP serving is now stateless for both modern and legacy clients, and Redis-backed sessions are gone, which is a real operational simplification for anyone self-hosting...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.