How to Prevent Prompt Injection in AI Agents
Keep untrusted pages, files, and tool output from changing an AI agent's permissions, destinations, or side effects.
Read articleAgentGuard Research
Security thinking for the agentic era.
Practical guidance, independent reviews, and clear explanations for teams building AI agents that can take real action.
How to Detect a Malicious MCP Server
Detect whether an MCP server is malicious from provenance, component behavior, metadata, telemetry, and target-state evidence, then contain it safely.
Read articleUnicode Prompt Attack Guide: Detect Invisible Instructions
Detect and control invisible or deceptive Unicode used to hide instructions in AI prompts and agent inputs.
Read articleHow to Prevent MCP Tool Poisoning
Prevent malicious instructions in MCP tool metadata from influencing an agent before or after installation.
Read articleHow to Scan an MCP Server Before Installation
A reproducible pre-installation MCP server scan process covering provenance, privileges, code, sandbox behavior, tool metadata, decision criteria, and regression.
Read articleMCP Authentication and Authorization Guide
Understand and design the authentication and authorization architecture for protected MCP deployments; not a step-by-step setup for one server.
Read articleHow to Secure MCP Authentication
Implement and verify authentication for one protected remote MCP server deployment.
Read articleA Practical Guide to AI Agent Security
Build an AI agent security program across threat modeling, component review, least privilege, runtime enforcement, evidence, and regression testing.
Read articleBrowser Agent Security: Risks, Controls, and Testing
Secure browser agents against prompt injection, unsafe navigation, credential exposure, consequential actions, and unverifiable outcomes.
Read articleHow to Prevent Credential Leaks from AI Agents
Protect AI agent credentials by mapping exposure paths, isolating secrets, brokering short-lived access, enforcing destinations, and testing response.
Read articleHow to Stop Malicious Tool Calls in AI Agents
Learn how to validate tool identity and arguments, enforce least privilege, gate high-impact actions, and verify AI agent tool-call outcomes.
Read articleCan You Trust an MCP Response? Four Failure Paths to Test
Prevent MCP response spoofing by separating endpoint identity, message integrity, prompt manipulation, and compromised-server behavior.
Read article