Works with the AI Agents you already use
Agent Tool-Safety Audit: Prompt-Injection Review
Audit LLM agent tools for prompt injection, unintended actions, and missing security guards.
$15
Agent Tool-Safety Audit: Prompt-Injection Review
Example session with this skill installed
Audit the security of my email-assistant agent tools in src/tools/ and the dispatcher in src/agent.py. Check for injection risks between reading emails and sending replies.
- Read your context and instructions
- Compiled the agent tool-safety audit
Audit complete. I have identified 3 high-severity security findings where untrusted email content can trigger unauthorized actions without a taint gate.
Agent Tool-Safety Audit Report
High Severity Findings
| Finding | Path | File:Line | Fix |
|---|---|---|---|
| Missing Taint Gate | read_inbox -> delete_email | src/tools/mail.py:42 | Implement a per-turn is_tainted flag. If read_inbox is called, set is_tainted = true. Block delete_email if is_tainted is true. |
| Model-Controlled Confirmation | send_reply accepts bypass_confirm | src/tools/mail.py:88 | Remove the bypass_confirm argument from the tool schema. Confirmation must be handled by the orchestrator via a fresh user turn. |
| Unattended Webhook Path | webhook_handler -> update_calendar | src/agent.py:112 | The webhook entry point lacks a human-in-the-loop gate. Require a secondary auth token or manual approval for calendar writes. |
Guard Check Summary
- Taint gate: FAILED. Acting tools are reachable in the same turn as untrusted reads.
- Fail-safe default: PASSED. New tools default to
actingstatus inBaseTool. - Sensitive-action confirmation: FAILED. The model can self-confirm sensitive actions via tool arguments.
- Ambiguity: PASSED.
get_contactrefuses and asks for clarification on multiple matches. - Honest results: PARTIAL. System prompt lacks definitions for
AMBIGUOUSandNEEDS_CONFIRMATIONerror strings. - Standing effects: PASSED. No scheduled task or memory writing tools detected.
Not Covered
- The
auth_providerimplementation was not provided; assumed to be secure for this audit.
Next steps
- Implement the
is_taintedstate logic insrc/agent.pyto bridge the gap between tool calls. - Update
src/tools/mail.pyto remove self-confirmation flags from the function signature. - Add the error string definitions to your system prompt to ensure the model handles refusals gracefully.
Connects securely to your tools. The creator never sees your data.
About this skill
Your LLM agents are one malicious email or webpage away from a breach. If an agent can read untrusted text and then call tools that act on the world, it is vulnerable to prompt injection and unintended actions. This skill audits your agent's tool definitions and execution flow to find and fix these critical security gaps.
What it does
- Inventory tools: Analyzes function schemas, MCP servers, and skill classes to map every available capability.
- Classify risk: Identifies "untrusted read" tools (emails, web, files) and "acting" tools (sending messages, moving money, controlling hardware).
- Audit guards: Checks for six specific security patterns including taint gates, fail-safe defaults, and sensitive-action confirmation.
- Trace attacker paths: Maps concrete exploit chains from untrusted input to unauthorized tool execution.
- Identify seams: Finds holes in unattended paths like cron jobs, webhooks, and tool-calling-tool routines.
- Report findings: Delivers a ranked list of vulnerabilities with file/line references and concrete code fixes.
How it works
- Tool mapping: Provide your agent's tool definitions, function schemas, or plugin manifests for review.
- Path analysis: The skill traces how data flows from external sources into acting tools.
- Guard verification: It compares your implementation against the
guards.mdsecurity patterns. - Fix generation: You receive a prioritized report following
report-template.mdwith specific code remediation steps.
Frameworks & tools
Works with any tool-calling architecture including LangChain, MCP servers, OpenAI function calling, plugin manifests, or custom Python/TypeScript tool loops.
Why this beats prompting it yourself
Generic prompts often miss "seams" like cron-job vulnerabilities or model-supplied confirmation flags. This skill uses a specialized methodology derived from high-stakes hardware control systems to systematically verify taint states and fail-safe defaults that general security prompts overlook.
Use cases
- Audit an MCP server before deploying it to a production environment.
- Review a customer service bot that reads inbound emails and updates a CRM.
- Harden a voice assistant that controls physical IoT devices or smart locks.
- Secure an autonomous researcher that browses the web and saves files to a local drive.
Known limitations
Cannot fully verify guards if the system prompt or orchestration loop code is missing. Reviews are limited to provided code fragments or stubs.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
Cannot verify guards if system prompts or orchestrator loops are missing. Audit is limited to provided tool definitions and code fragments.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed