skills.rest
nguyenpv1980-wq
nguyenpv1980-wqCOMMUNITY
Explore all skills by this author →
GitHub

agent-tool-safety-guard

Review AI agent tool and API access for safety boundaries.

2|Updated Jul 6, 2026
Search Tags:#approval gates#least privilege#blast radius#agent safety#tool permissions#identity binding
One-click install
npx skills add https://github.com/nguyenpv1980-wq/Project-Aegis --skill agent-tool-safety-guard
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: agent-tool-safety-guard
Source: https://github.com/nguyenpv1980-wq/Project-Aegis/tree/main/.claude/skills/agent-tool-safety-guard
Command: npx skills add https://github.com/nguyenpv1980-wq/Project-Aegis --skill agent-tool-safety-guard

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you prevent an AI agent from having too much power, too much access, or the wrong identity when it calls tools, functions, or APIs. It is designed to stop confused-deputy failures, unsafe destructive actions, and tool-chain abuse before they become incidents.

Core Features & Use Cases

  • Least-Privilege Tool Review: Builds a per-tool matrix covering side effects, reversibility, blast radius, cost, and execution identity.
  • Calling-User Authority Binding: Ensures each tool runs with the calling user's permissions and tenant scope, not a shared privileged service account.
  • Safety Controls for High-Risk Actions: Requires schema validation, approval gates for destructive or expensive actions, and limits on tool-chain composition abuse.
  • Use Case: Review a support agent that can look up customers, issue refunds, and delete accounts, then narrow unsafe tools, add approvals, and define safe execution boundaries.

Quick Start

Ask for a tool-safety review of the agent's available tools, including privilege boundaries, approval requirements, and misuse paths.

Frequently Asked Questions about agent-tool-safety-guard

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I prevent an AI agent from running destructive API calls with a shared privileged service account?▼

To prevent destructive API calls, enforce calling-user identity binding so each tool runs with the caller's permissions and tenant scope rather than a shared privileged account. This stops confused-deputy failures by ensuring the agent never inherits blanket service-account authority.

What is the best way to review agent tool permissions and limit blast radius for high-risk actions?▼

The best way to limit blast radius is building a per-tool review matrix covering side effects, reversibility, cost, and execution identity. High-risk destructive or expensive actions then get schema validation and approval gates before execution.

How do approval gates work for AI agent tools that can spend money or delete records?▼

Approval gates for agent tools intercept expensive or destructive actions before execution, requiring explicit authorization. They pair with argument validation to ensure the tool inputs match expected schemas, preventing unauthorized spending or record deletion.

Can I use this tool-safety review for a support agent that issues refunds and deletes customer accounts?▼

Yes, you can review a support agent that looks up customers, issues refunds, and deletes accounts. The review narrows unsafe tools, adds approval requirements for destructive actions, and defines safe execution boundaries to prevent misuse.

When do I need tool-chain composition abuse analysis for my AI agent's API access?▼

You need tool-chain composition abuse analysis when an agent can chain tool outputs into further actions. This analysis identifies misuse paths where combining seemingly safe individual tools creates unintended destructive or unauthorized cascading behavior.

Why does my AI agent need per-tool permission mapping instead of a single global access policy?▼

Per-tool permission mapping is needed because individual tools have different side effects, reversibility, and blast radius. A single global policy over-privileges the agent, whereas mapping each tool enforces least-privilege boundaries and isolates high-risk actions.

Related Skills

Explore More Software Engineering Skills→
viktorbezdekviktorbezdek
10

trade-off-analysis

Analyze product option trade-offs and produce a trade-off matrix with costs and kill conditions.

Community
Intermediate
rauly-1408rauly-1408

dag-scope-enforcer

Enforce runtime permission boundaries for AI agent tool, file, command, and network operations.

Community
Advanced
invowkinvowk

invowk-schema

Enforce consistent schemas for invkfile.cue and invkmod module definitions.

Official
Advanced
jaskaranhundaljaskaranhundal
3

identity-access-risk

Analyze AWS IAM anomalies and privilege escalation attempts from CloudTrail logs.

Community
Advanced