Generative AI Guardrails
Generative AI guardrails (GenAI Guardrails) are deterministic checks that run before a generative AI agent executes a tool call. SigmaShake is that product at the tool-call boundary: a local policy engine that evaluates every action, then ALLOW, DENY, ASK, or FORCE a safe alternative in under 2 milliseconds, with no LLM in the enforcement loop.
Also spelled Gen AI Guardrails. Agent-specific forms — Generative AI Agent Guardrails and GenAI Agent Guardrails — are the same product: SigmaShake at the tool-call boundary. Related but distinct from AI agent guardrails: that URL owns that phrase; this URL owns the generative-AI / GenAI search terms. Same engine.
The same product covers tool and action guardrails — AI tool guardrails and AI action guardrails — at sigmashake.com. Building generative AI guardrails is this page, not a separate product: honest-agent policy at the tool-call boundary, not an LLM text filter and not a sandbox.
What generative AI guardrails actually gate
Coding agents such as Claude Code, Cursor, and Codex can read files, write files, run shell, and call MCP tools. A system prompt is a probabilistic nudge. Generative AI guardrails are the enforcement path: a typed IF/THEN rule set evaluated before the action runs. See the four threats on AI agent security.
Not an LLM text filter. Not a sandbox. Not the book.
NVIDIA NeMo Guardrails, Guardrails AI, OpenAI moderation, and IBM/McKinsey explainers inspect generated text. SigmaShake does not score generated text. It decides whether the tool call is allowed, before Bash, file writes, or MCP tools run. It is also not a sandbox: SigmaShake is honest-agent / integration-dependent. For adversarial isolation, compose with OS-level sandboxing (Docker, seccomp, or Apple Sandbox). The book AI Agent Guardrails by Charles Chong is a separate $5 one-time purchase, not this product.
30 seconds to protected
1. Install
curl -fsSL install.sigmashake.com | sh2. Wire it in
ssg initReview the Terms of Service. See the docs. Free to download and try; $5/month unlocks Pro value surfaces.
FAQ
What are generative AI guardrails?
Deterministic checks that run before a generative AI agent executes a tool call. They ALLOW, DENY, ASK, or FORCE a safe alternative instead of hoping the model follows instructions. SigmaShake is that product at the tool-call boundary, local, under 2 milliseconds, no LLM in the loop.
What are GenAI guardrails?
The same product. GenAI Guardrails (also spelled Gen AI Guardrails) is SigmaShake at the tool-call boundary: local policy before Bash, file, or MCP tools run.
What are Generative AI Agent Guardrails?
The same product. Generative AI Agent Guardrails is SigmaShake at the tool-call boundary — the agent-specific form of the generative AI guardrails search.
What are GenAI Agent Guardrails?
The same product. GenAI Agent Guardrails is SigmaShake at the tool-call boundary — the agent-specific form of GenAI Guardrails.
What else are they called?
Tool and action guardrails, AI tool guardrails, and AI action guardrails. Building generative AI guardrails (or building GenAI agent guardrails) is the same SigmaShake product at the tool-call boundary.
How is this different from NeMo, Guardrails AI, or OpenAI?
Those inspect generated text. SigmaShake gates the tool call before Bash, file writes, or MCP tools run.
Do generative AI guardrails stop an agent from running rm -rf?
Yes, when a matching rule is installed: DENY or FORCE before the shell runs. Honest agent, not a sandbox. See stop rm -rf.
Same as AI agent guardrails?
Related phrase, same engine. This page is the gen-AI search term. The sister page is AI agent guardrails.
How fast are they?
Typical local evaluation is under 2 milliseconds. There is no LLM call in the enforcement path.
Do they work with Claude Code and Cursor?
Yes. Claude Code, Codex, Cursor, VS Code Copilot, Antigravity, Aider, OpenCode, and any MCP-compatible agent. Evaluation stays on your machine.