How does this tool differ from traditional web application security scanners?

This skill specifically addresses the unique vulnerabilities of LLM architectures, such as recursive prompt injection and vector database isolation, which generic cloud security scanners typically overlook.

Which AI agent frameworks or RAG platforms are compatible with this security review tool?

It is compatible with any agentic framework or RAG pipeline, provided you can describe the architecture (data sources, tools, and permissions) in a standard YAML or JSON format.

How are new LLM-specific vulnerabilities and OWASP updates integrated into the tool?

The skill is updated periodically to reflect the latest evolving threats and official updates to the OWASP Top 10 for LLM Applications (2025) list.

What exactly is included in the purchase of this security review skill?

You will receive the specialized Python-based CLI tool, the structured questionnaire framework, and the Markdown reporting templates required to perform a full audit.

Do I need to grant the tool live access to my production environment to perform a threat model?

No, this tool acts as a static architecture and logic analyzer; it does not require access to your live production environment or active API keys to identify potential exploit paths.

Can this tool help identify multi-tenant data leakage risks in a shared RAG system?

Yes, the tool is designed to identify flaws in tenant isolation and retrieval logic, helping you prevent cross-tenant data leakage in shared vector databases.

LLM Application Security Review

Example session with this skill installed

Audit my RAG agent architecture. It pulls from a public web crawler and can send emails on behalf of the user. I'm worried about malicious site content triggering unauthorized emails.

Read your context and instructions
Compiled the llm application security

I have analyzed your RAG agent architecture focusing on the path from the web crawler (untrusted input) to the email tool (high-impact action). This setup is highly vulnerable to Indirect Prompt Injection.

LLM Security Architecture Review

Risk ID	Vulnerability	Severity	Description
LLM01	Indirect Prompt Injection	Critical	Malicious content from crawled sites can hijack the model context to send unauthorized emails.
LLM02	Excessive Agency	High	The email tool appears to lack a "human-in-the-loop" requirement for high-impact actions.
LLM06	Sensitive Data Disclosure	Medium	Model output rendered directly into emails could leak system instructions or internal metadata.
LLM10	Unbounded Consumption	Low	A recursive crawl could lead to token exhaustion and increased API costs.

Remediation Plan

Implement Human-in-the-Loop: Require a cryptographic signature or manual UI approval before the email tool executes a send command.
2.

Input Sanitization: Pass crawled data through a dedicated llm-prompt-injection-scanner before it enters the RAG context.
3.

Least Privilege: Restrict the email tool's scope to specific domains or pre-approved templates to prevent arbitrary content delivery.
4.

Output Encoding: Ensure all model-generated text is treated as untrusted and encoded before being inserted into the email body.

Next steps

Generate the full arch.yaml to map all tool permissions.
Run the questionnaire command to evaluate your session management and tenant isolation.
Integrate a prompt injection filter on the ingestion pipeline.

Connects securely to your tools. The creator never sees your data.

About this skill

LLM applications introduce unique attack vectors like prompt injection and insecure output handling that traditional scanners miss. This skill provides a specialized security review framework for RAG systems and AI agents, mapping system architectures against the OWASP Top 10 for LLM Applications (2025).

What it does

Architecture analysis identifies dangerous data paths from untrusted sources to high-impact tools.
Threat modeling generates structured reports based on YAML or JSON system descriptions.
Compliance mapping aligns application logic with the OWASP Top 10 for LLM vulnerabilities.
Vulnerability scoring evaluates risks like excessive agency, system prompt leakage, and cross-tenant retrieval.
Remediation guidance provides specific fixes for model poisoning and unbounded token consumption.

How it works

Define architecture by creating a YAML file describing data sources, tools, and permissions.
Run analysis using the arch command to map potential exploit paths in the agent workflow.
Conduct questionnaire to assess operational security controls and scoring.
Generate report to export a Markdown threat model with prioritized remediation steps.

Frameworks & tools

This tool targets systems built with RAG pipelines, MCP servers, and tool-using agents. It utilizes a Python-based CLI for analysis and reporting.

Why this beats prompting it yourself

Generic security prompts often miss architectural flaws like improper output encoding or vector database tenant isolation. This skill uses a systematic questionnaire and automated path analysis to ensure no OWASP LLM category is overlooked.

Use cases

Auditing a RAG pipeline for sensitive information disclosure risks.
Reviewing an autonomous agent's tool permissions to prevent excessive agency.
Hardening a customer-facing chatbot against prompt injection and system leakage.
Validating multi-tenant isolation in vector database retrieval logic.

Known limitations

Requires a manual YAML/JSON description of the application architecture to perform analysis. Scoring depends on accurate user input during the questionnaire phase.

How to use

Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.

Reviews

No reviews yet

Be one of the first to try it. Every listed skill passes our trust checks below.

Security scanned

Passed our 8-point scan before listing

Fresh listing

Recently published to Agensi

30-day refund

Not a fit? Get your money back

LLM Application Security Review