AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
786 articles
MCP server security for AI coding agents
A critical RCE in Anthropic's own MCP Inspector (CVSS 9.4) and two Cursor CVEs show that giving AI agents tool access creates a new, largely unvetted attack surface.
A Reproducible Rubric for Measuring Prompt-Injection Risk in Agent Skills
OWASP has ranked prompt injection the #1 LLM risk for two straight editions, yet almost no one scores agent skill packages for it consistently. Here's a rubric.
Scanning AI-Generated Code Before It Merges: Wiring Scanners into Coding Assistants with MCP
Research found ~40% of Copilot suggestions were vulnerable, and devs using AI assistants trusted their code more. MCP lets you scan before merge.
Auditing AI agent skill registries for hardcoded keys
29M new hardcoded secrets hit public GitHub in 2025, up 34% YoY — and 3% of MCP servers in production carry hardcoded credentials as theft traps.
Securing MCP Servers for AI Agents
Five CVEs in 2025 alone trace MCP tool compromise back to one root cause: unsanitized strings piped into exec(). Here's how to expose and consume MCP safely.
Signing and provenance standards for AI agent skill registries
Shai-Hulud infected 500+ npm packages via stolen tokens in 2025. Agent skill registries are repeating the same unsigned-artifact mistake — here's the fix.
Threat-modeling for AI-native applications
STRIDE has six categories from 1999. OWASP's LLM Top 10 and MITRE ATLAS's ~84 techniques show why agentic AI needs new threat-modeling columns, not a new framework.
The OWASP Top 10 for LLM Applications, Explained
The OWASP LLM Top 10 is the closest thing the field has to a shared checklist for AI security. Here is what each of the ten risks actually means, in plain language, with the defenses that matter.
MCP Protocol News: The 2026 Spec Update and Its Security Fallout
The biggest MCP protocol news of 2026 is a major specification overhaul that fixes old flaws and shifts a load of new security responsibility onto developers.
Red-Teaming AI Applications: A Field Guide
You cannot secure an LLM application by reading its code alone. You have to attack it the way an adversary will — with language, with poisoned content, and with the goal of making it do something it should not. Here is how to run an AI red team.
What Is the Model Context Protocol (MCP)? And What It Means for Security
MCP is the USB-C of AI integrations — one open standard for connecting models to tools and data. It also standardizes a fresh attack surface, so understanding both halves matters.
Agents Can Now Procure Safeguard Through MCP
AI agents can browse regional pricing, compare tiers, start a Stripe checkout, and verify activation — the entire Safeguard procurement journey now runs through the MCP server.
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.