Works with the AI Agents you already use
agent attack surface review
Audits AI agent configurations and MCP servers for security risks, secret exposure, and tool overreach.
$15
agent attack surface review
Example session with this skill installed
Audit my current Claude Code and MCP setup for security risks. The agent is intended for local refactoring only. Check mcp.json and my project rules for any overreach.
- Read your context and instructions
- Compiled the agent attack surface
I have completed the security review for your Claude Code setup. The audit focused on your project root and the local mcp.json configuration.
Agent Attack Surface Review
Scope
- Agent under review: Claude Code
- Boundary: Project root (
/workspaces/project-alpha) and~/Library/Application Support/Claude/config.json - Intent: Local code refactoring and linting.
- Exclusions: External CI/CD pipelines and cloud provider IAM roles.
Verdict: ACCEPTABLE WITH FIXES
The agent has broad read/write access to the local filesystem but lacks an approval gate for shell execution.
Surface Summary
- Tools/MCP Servers: 4 (Filesystem, Shell, Google Search, SequentialThinking)
- Ingress Channels: 2 (Terminal input, Web search results)
- Approval Gates: 1 (Write-access only)
Findings
- AS-04 | Unsupervised high-impact action | Severity: High
- Evidence:
mcp.json:12-"shell": {"allow_all": true} - Impact: Untrusted input from web search could trigger arbitrary command execution without a prompt.
- Fix: Change to an allowlist of specific commands or enable
auto_approve: falsefor the shell tool.
- Evidence:
- AS-01 | Secret exposure | Severity: Medium
- Evidence:
.cursor/rules/main.md:45-export GITHUB_TOKEN=ghp_a1b2... - Impact: The token is visible in the agent's context window and could be exfiltrated.
- Fix: Move the token to a
.envfile and add it to.gitignore.
- Evidence:
Patch Plan
- P0: Disable
allow_allinmcp.jsonand move the GitHub token to environment variables. - P1: Add an explicit tool allowlist to the agent instructions.
Unverified
- I could not verify if the
SequentialThinkingtool has internal logging enabled. Check the server source code for audit trail implementation.
{
"agent": "claude-code",
"verdict": "ACCEPTABLE WITH FIXES",
"critical_findings": 0,
"high_findings": 1,
"remediation_status": "pending"
}
Next steps
- Review the
mcp.jsonfile to restrict shell tool permissions. - Run
scripts/scan_surfaces.pyagain after moving your secrets to verify redaction. - Request a remediation diff if you want me to generate the updated configuration files.
Connects securely to your tools. The creator never sees your data.
About this skill
AI agents with autonomous tool access introduce significant security risks, from secret exposure to unsupervised high-impact actions. This skill audits your agent configurations to identify where untrusted input can reach sensitive tools or credentials. It provides a structured, evidence-backed security review for coding agents, MCP servers, and custom agentic runtimes.
What it does
- Inventory surfaces by mapping all tools, skills, and MCP servers the agent can invoke and their specific capabilities.
- Trace reach to identify every file, repository, and database the agent can touch through its toolset.
- Identify ingress paths where untrusted content like web pages or PR comments could trigger prompt injection.
- Audit controls by checking for human approval gates, sandboxes, rate limits, and logging.
- Rank risks across eight security classes including tool overreach and data exfiltration.
How it works
- Define scope by specifying the agent runtime, project root, and the intended purpose of the agent.
- Run inventory using the internal scanner to map config surfaces, authority, and reach.
- Analyze risks against the severity rubric, checking for unapproved high-impact paths.
- Generate report including a security verdict, evidence-backed findings, and a prioritized patch plan.
Frameworks & tools
Works with Claude Code, Cursor, Codex CLI, OpenClaw, OpenCode, and custom agent runtimes. Audits mcp.json, CLAUDE.md, AGENTS.md, .cursor/rules, and SKILL.md configurations.
Why this beats prompting it yourself
Generic security prompts often miss hidden execution paths in nested MCP configurations or fail to map how ingress data flows into tools. This skill uses a systematic eight-class risk catalog and a dedicated scanner to ensure no surface is skipped. It enforces evidence rules that require specific file paths and line numbers, preventing vague or hallucinated security claims.
Use cases
- Pre-shipment audit of a new coding agent before allowing it access to a private repository.
- MCP hardening to ensure a local server doesn't provide unverified network or shell access.
- Credential hygiene to find and redact secrets accidentally committed to agent instruction files.
- Permission review when upgrading an agent's authority to include write or spend capabilities.
Known limitations
Does not perform live penetration testing, external host scanning, or compliance certification. Audit is strictly read-only and limited to the local project root and user-level MCP configs.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.
Read-only audit of local config files and project root only. No live penetration testing or external network scanning. Scope limited to provided project files and MCP settings.
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 23 days ago
- Software
- @scope/server
Node.js
Permissions
Read-only. The skill reads agent config and instruction files inside the declared scope, then reports findings. It does not modify files, run project code, install packages, or send data anywhere. The optional scanner also only reads files and redacts secret values in its output.