skills.rest
FluxonLab
FluxonLabCOMMUNITY
Explore all skills by this author →
Website𝕏@fluxonlabGitHub

ai-security-review

Reviews LLM applications for prompt injection, tool misuse, exfiltration, and delegation vulnerabilities.

2|Updated Jun 1, 2026
Search Tags:#mcp#security review#red team#prompt injection#data exfiltration#agent security#llm security
One-click install
npx skills add https://github.com/FluxonLab/Skillry --skill ai-security-review-fluxonlab
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: ai-security-review
Source: https://github.com/FluxonLab/Skillry/tree/main/plugins/security/skills/48-ai-security-review
Command: npx skills add https://github.com/FluxonLab/Skillry --skill ai-security-review-fluxonlab

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? LLM-based applications introduce an attack surface that traditional security reviews miss: prompt injection, tool misuse, data exfiltration through tool arguments, and agent delegation abuse. This Skill provides a structured procedure to test and document these AI-specific vulnerabilities before production exposure. ## Core Features & Use Cases - Prompt Injection Testing: Tests direct injection patterns and indirect injection via retrieved web pages, PDFs, and database records, with documented success rates. - Tool and Exfiltration Audits: Constructs adversarial scenarios for every agent tool and tests each exfiltration channel with synthetic canary strings. - Governance and MCP Gatekeeping: Includes merged references for agent permission audits (least privilege, approval gates, audit logging) and for deciding whether to activate external MCP connectors with minimum scope. - Use Case: Before launching a customer-facing agent with web browsing and email tools, run this review to map the attack surface, attempt indirect injection through a test web page, verify output sanitization, and produce a prioritized finding report. ## Quick Start Use the ai-security-review skill to perform a full AI security review of my agent application before it goes to production.

Frequently Asked Questions about ai-security-review

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

FAQPage Schema
How do I test an LLM application for prompt injection?▼

Test direct injection with patterns like instruction overrides, fabricated system messages, and multi-turn role-play shifts, documenting success rates across 5 runs. Then test indirect injection by embedding hidden instructions in web pages, PDFs, or database records the agent retrieves.

What is indirect prompt injection and why does it matter?▼

Indirect prompt injection places malicious instructions in external content the agent retrieves, such as web pages or uploaded files, rather than in the user's message. It is the highest-impact vector for agents with web access because any retrieved content is attacker-controlled.

When should I run an AI security review?▼

Run one before production launch, after a security incident involving an AI component, when new tools are added to an agent, or when the system processes untrusted content. Traditional OWASP reviews do not cover the AI-specific attack surface.

Can a system prompt prevent prompt injection attacks?▼

No. A system prompt is a preference expression, not a security control, and adversarial inputs will override it in some percentage of cases. Structural controls must be enforced in application code: tool argument validation, output filtering, and permission checks.

What are the limitations of this AI security review approach?▼

It covers only the AI-specific attack surface, not traditional vulnerabilities like OWASP Top 10, and excludes component provenance concerns handled by supply-chain review. Purely read-only LLMs with no tools or external data ingestion have minimal applicable surface.

How do I safely test data exfiltration in an AI agent?▼

Use clearly synthetic canary strings like CANARY_TEST_20250531_DO_NOT_USE, never real credentials or internal hostnames. Test each channel (HTTP, email, webhooks, file writes) in a dedicated non-production environment and confirm blocking is code-level, not prompt-level.

Related Skills

Explore More Software Engineering Skills→
nguyenpv1980-wqnguyenpv1980-wq
2

threat-modeler

Build security threat models with STRIDE analysis and mitigations.

Community
Advanced
QuyDuQuyDu

audit-plan-remediation

Convert validated audit findings into prioritized remediation plans with owners, dependencies, and rollback guidance.

Community
Advanced
OWASPOWASP
141

mcp-server-review

Audits MCP server implementations for security risks across seven defined review areas.

Official
Intermediate
ArianHobson333ArianHobson333
7

ai-threat-testing

Detect and exploit OWASP LLM Top 10 vulnerabilities in large language models.

Community
Advanced