PrismorSec/prismor

What it solves

Prismor provides runtime security for AI agents (such as Claude Code, Codex, and LangChain) to prevent risks associated with autonomous execution. It addresses threats like prompt injection, unintended destructive actions, secret exfiltration, privilege escalation, and supply chain risks that traditional OS-level security tools often miss because they lack AI-specific context.

How it works

Prismor intercepts tool calls from AI agents through various integration surfaces—including shell hooks, an LLM proxy, and a dedicated MCP gateway. Every proposed action is evaluated against a customizable policy engine before execution. The system can operate in "observe" mode (logging activity for auditing) or "enforce" mode (blocking dangerous actions in real-time). It also includes a "Semantic Guard" for LLM-assisted intent checks to catch paraphrased injections that regex rules might miss.

Who it’s for

It is designed for developers and organizations deploying AI agents that have the ability to execute shell commands, read/write files, and call external APIs, particularly those operating in regulated environments or handling sensitive data.

Highlights

• Multi-Agent Support: Integrates with a wide array of agents including Claude Code, Cursor, GitHub Copilot, and various framework agents like LangChain and CrewAI.
• MCP Guardrails: Provides a gateway to block specific Model Context Protocol (MCP) servers or tools and scans tool responses for injections.
• Secret Protection: Features "Sweep and Cloak" to prevent secrets from crossing tool boundaries and uses honeytoken canaries to detect reconnaissance.
• Tamper-Evident Auditing: Hash-chains and signs agent actions locally to create a verifiable audit trail.
• Governance Modes: Offers pre-configured security postures (e.g., dev-safe, regulated-airgap) to balance friction and security.
• LLM Proxy: Governs agents that cannot be hooked directly by intercepting model traffic and judging proposed tool calls.