AgentShield

•
4
•

Intercepts package installation requests from AI agents, checking them against CVEs, static analysis, and enforcing security policies to prevent supply chain attacks.

AI agents now have the capability to install arbitrary software, introducing significant security risks such as prompt injection, typosquatting, and the installation of backdoored or vulnerable packages. AgentShield acts as a crucial security layer, sitting directly between an agent's intent to install a package and the system's execution of that request. It intercepts package installations across multiple ecosystems (pip, npm, cargo, gem, go) through various integration points, performing real-time checks against CVE databases, identifying typosquats and malicious packages, and enforcing customizable security policies, all while ensuring local operation and no telemetry. This comprehensive approach safeguards AI agents from supply chain attacks, protects sensitive data like API keys and context windows, and generates detailed security posture reports.

主要功能

012 GitHub stars
02Configurable security policies to allow, block, or warn on detected risks
03Comprehensive vulnerability scanning against CVE databases (OSV, NVD, GitHub Advisory)
04Local-first operation with offline mirror capabilities and no telemetry
05Real-time interception of package installations for various ecosystems (pip, npm, cargo, gem, go)
06Detection of typosquatting, malicious packages, and prompt-injected install attempts

使用案例

01Auditing package installation requests made by AI agent frameworks against CVE databases and static analysis
02Generating Security Bill of Materials (SBOMs) and security posture reports for agent dependencies
03Protecting AI agents from installing malicious or vulnerable packages through prompt injection or typosquatting