AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
786 articles
Prompt injection attacks: direct vs indirect
Direct prompt injection comes from the chat box; indirect injection hides in the data your AI agent trusts. Here's how the two attack types differ and what stops each.
Agent hijacking: the real-world impact of prompt injection
From a zero-click Microsoft 365 Copilot breach to poisoned MCP servers, AI agent hijacking is now a real, documented software supply chain threat.
Securing Model Context Protocol (MCP) servers
MCP server security explained through real 2025 CVEs, tool poisoning, and rug-pull attacks, plus concrete controls security teams need to defend AI agent tool calls.
mcp-scan: detecting malicious MCP tool definitions
MCP lets AI agents call tools via plain-text descriptions the model trusts blindly. Here's how mcp-scan catches poisoning, rug-pulls, and shadowing.
Can AI write secure code? Auditing AI-generated code
AI writes code fast, but studies from 2021 to 2025 show it also reproduces insecure patterns and invents fake dependencies. Here's what the data says.
GitHub Copilot code security: XSS vulnerabilities found in React
Copilot commonly suggests dangerouslySetInnerHTML and unsanitized DOM writes in React. Here's the data on AI-generated XSS risk and how to catch it.
How Copilot amplifies insecure codebases
Copilot writes ~46% of code where enabled, and studies show ~40% of its security-relevant suggestions are vulnerable. Here's the data on the risk.
How to Audit the Dependencies of an AI Agent
An AI agent's dependency tree spans packages, MCP servers, models, and system prompts. A step-by-step audit method that actually enumerates all four layers.
AI-Generated Code Security: risks and controls
AI now writes up to 40%+ of new code, and models hallucinate nonexistent packages in 5-22% of outputs. Here's why Black Duck-style SCA misses that risk, and what controls actually work.
5 best practices for adopting GitHub Copilot securely
GitHub Copilot has 1.3M+ paid seats. Five concrete, evidence-based practices for locking down content exclusion, licensing, code quality, and prompt injection risk.
AI hallucinations and their security implications for developers
LLMs hallucinate nonexistent packages in up to 1 in 5 code samples — and slopsquatting attacks are already exploiting that predictability in the wild.
GenAI Code Review Tools: A 2025 Field Test
We field-tested five GenAI code review tools against 240 seeded security defects to see which catch real issues and which hallucinate findings.
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.