Works with the AI Agents you already use
agent shield
Perform evidence-based security audits and risk scoring for AI agents, MCP tools, and system prompts.
Free
agent shield
Example session with this skill installed
Perform a security preflight on this MCP tool definition for a local file-search agent. It has read access to the home directory and uses a recursive search function.
- Read your context and instructions
- Compiled the agent shield
I have completed the security preflight for the file-search MCP tool.
Executive Verdict
- Security Score: 72/100
- Overall Risk: High
- Audit Confidence: 85%
- Recommendation: Do not deploy without restricting the root search path.
Security Findings
| ID | Severity | Finding | Impact | Penalty |
|---|---|---|---|---|
| AS-001 | High | Excessive Directory Permissions | Sensitive file exposure | 7 |
| AS-002 | Medium | Runaway Recursion | Resource exhaustion/DDoS | 3 |
| AS-003 | Low | Missing Search Boundaries | Information leakage | 1 |
Failure Path
- Injection: An untrusted prompt asks the agent to search for
.ssh/id_rsa. - Execution: The agent uses the
recursive_searchtool starting from the home directory. - Exfiltration: The tool successfully reads and returns private key contents to the LLM context.
Fix Pack
- Current Problem: Tool allows searching from
~/which includes sensitive credentials. - Recommended Change: Hardcode the search root to a specific project directory or use an allowlist.
- Safe Example:
const SEARCH_ROOT = '/home/user/projects/current';
Next steps
- Implement a search directory allowlist to prevent access to hidden system folders.
- Add an action limit or depth cap to the recursive search function to prevent runaway loops.
- Review the
Residual Risksection regarding missing network egress logs.
Score Calculation
Starting score: 100
Finding penalties: -11
Control bonus: +0
Final Security Score: 72/100
Audit Confidence: 85%
Connects securely to your tools. The creator never sees your data.
About this skill
AI agents often ship with over-privileged tools, hidden prompt injection vectors, and runaway autonomy that leads to high compute costs or data leaks. Agent Shield provides an evidence-based security preflight for AI agents, MCP configurations, and agentic workflows. It identifies critical weaknesses before you deploy or publish.
What it does
- Vulnerability Analysis checks for prompt injection, goal hijacking, and tool-description poisoning.
- Permission Auditing evaluates excessive agency against an intended capability baseline.
- Supply Chain Review inspects MCP server definitions, tool shadowing, and dependency risks.
- Risk Scoring generates a calibrated security score (0-100) based on impact, exploitability, and controls.
- Remediation Planning provides specific failure paths, recommended fixes, and verification steps.
How it works
- Analyze Material by reviewing the agent's instructions, tool definitions, and permission policies.
- Establish Baseline by determining what the agent actually needs to accomplish its purpose.
- Execute Domains across 10 security categories including memory, network, and autonomy.
- Calculate Scores using a mathematical penalty model and audit confidence metric.
Frameworks & tools
Works with any agentic architecture including Claude Code, Cursor rules, MCP (Model Context Protocol) servers, and custom tool-calling agents.
Why this beats prompting it yourself
Generic security prompts often hallucinate vulnerabilities or miss the nuance of agent autonomy. Agent Shield uses a strict evidence standard and a mathematical scoring system to ensure findings are objective, deduplicated, and actionable.
Use cases
- Pre-deployment Audit of new MCP tools or agentic skills before company-wide rollout.
- Prompt Hardening to identify injection vectors in complex multi-step instructions.
- Permission Scoping to ensure coding agents have appropriate filesystem and network access.
- Cost Control Review to detect potential runaway loops and unbounded tool execution.
Known limitations
This is a static security review aid, not a runtime penetration test or a guarantee of safety. It cannot verify implementation details that are not provided in the supplied documentation or code.
Static assessment only. No runtime testing. Does not certify compliance. Results depend entirely on the quality of supplied material.
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 any payment.
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
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 28 days ago
- Free with an Agensi account