PrismorSec/prismor

What it solves

Prismor provides runtime security for AI coding agents (such as Claude Code, Codex, and Cursor) and production AI frameworks. It addresses risks that traditional security tools miss, such as prompt injection, unintended destructive actions, secret exfiltration, privilege escalation, and supply chain vulnerabilities introduced by autonomous agents executing shell commands or accessing files.

How it works

Prismor acts as a security layer between the AI agent and the system. It intercepts tool calls through runtime hooks, an MCP gateway, or framework adapters. Every action is evaluated against a customizable YAML-based policy engine before execution. Depending on the rule, Prismor can either observe (log the action) or enforce (block the action in real time).

Who it’s for

It is designed for developers and organizations using autonomous AI coding agents or building production-grade agentic workflows who need to ensure these agents operate within safe, least-privilege boundaries.

Highlights

• Multi-Agent Support: Integrates with a wide range of coding agents (Claude Code, Cursor, GitHub Copilot, etc.) and frameworks (LangChain, CrewAI, Pydantic AI, etc.).
• MCP Gateway: A single connector that evaluates every tools/call and scans responses for prompt injection before they reach the model.
• Secret Protection: Features "Sweep and Cloak" to prevent secrets from being leaked at tool boundaries and canary honeytokens to detect reconnaissance.
• Tamper-Evident Auditing: Uses hash-chains and Ed25519 signatures to create a signed audit trail of all agent actions.
• Flexible Enforcement: Offers per-rule observe/enforce modes and layered policies (org, project, and repo levels) with a non-overridable safety floor.
• Semantic Guard: An optional LLM-assisted layer to detect paraphrased prompt-injection attempts that bypass regex rules.