AI Security
MCP tool annotations are hints, not guarantees
The MCP spec says clients must treat tool annotations as untrusted unless the server is trusted. What that means for approval logic and auto-run rules.
Oct 18, 20263 min read
Resources
Supply Chain Security, in plain English.
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
Filter
Articles
RSS feedAI Security
MCP token theft: short-lived tokens and refresh rotation
Stolen MCP tokens look like legitimate requests. The spec asks for secure storage, short-lived access tokens and refresh rotation. Details and a checklist.
Oct 18, 20263 min read
AI Security
MCP token audience and resource indicators, explained simply
How the resource parameter from RFC 8707 binds an MCP access token to one server, and what an MCP server must check before trusting a token.
Oct 18, 20263 min read
AI Security
Why an MCP session ID is not authentication
MCP servers must not use sessions for authentication. Learn the two session hijack attacks and how to bind session IDs to the authenticated user.
Oct 18, 20263 min read
AI Security
What an MCP server must do on every tool call
The MCP tools spec says servers must validate inputs, enforce access control, rate limit and sanitize outputs. A practical checklist for building a tool.
Oct 18, 20264 min read
AI Security
MCP HTTP servers: Origin checks and DNS rebinding explained
Local MCP servers over HTTP can be reached from a website through DNS rebinding. The spec requires Origin validation and recommends binding to localhost.
Oct 18, 20263 min read
AI Security
MCP confused deputy: how proxy servers skip user consent
An MCP proxy with a static client ID and a consent cookie can hand an attacker an authorization code. Learn the four conditions and the required fixes.
Oct 18, 20263 min read
AI Security
MCP client SSRF: defending OAuth discovery from hostile servers
A malicious MCP server can point OAuth discovery URLs at internal services and cloud metadata. The spec lists the defenses clients should apply.
Oct 18, 20263 min read
AI Security
MCP clients: validate authorization URLs to stop XSS and RCE
A hostile MCP server can send a javascript: or shell-laced authorization URL. Clients must allow only http and https and never open URLs through a shell.
Oct 18, 20263 min read
Stay informed
Weekly insights on software supply chain security, delivered to your inbox.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.