High-impact vulnerabilities confirmed in real AI package ecosystems.
Scan or Search
Run a free AI security audit on any GitHub repo, or search our database
Free AI-powered security audit3 scans/dayPublic repos only
How It Works
Add the skill, scan packages, get trust scores
1. Add Skill
2. Security Scan
analyzing code...
1
import { exec } from "child_process"2
const API_KEY = process.env.KEY3
exec(`curl ${userInput}`) // ⚠️4
fs.readFileSync("/etc/passwd")5
return { status: "ok" }3. Trust Score
Safe (80+)
Review (50-79)
Risky (<50)
Real Threats We've Found
Live DataNot theoretical risks. Actual malicious code discovered in AI packages.
193
Packages Scanned
and counting...
Most Common
Credential Theft
Packages stealing API keys and environment variables.
process.env.OPENAI_API_KEY Active
31
Critical Threats
Data Exfiltration
Sending your code to external servers.
Prompt Injection
Hidden instructions manipulating agents.
303
Total Findings
Obfuscation
Hidden malicious payloads.
337
Audits Completed
Live Security Feed
Real vulnerabilities discovered in AI packages — updated in real-time
✍️ Latest Research
Featured investigation plus fast-scan briefs. Structured like an editorial desk, not a cloned card wall.
Featured Investigation
Context Pivoting: A New Attack Vector in Multi-Server MCP Deployments
How a single malicious MCP server can hijack your entire agent — without touching the other servers.
02
How Multi-Agent Consensus Makes Security Audits More Reliable
Why multiple independent AI agents auditing the same package — then cross-validating — beats any single scanner.
03
The State of MCP Server Security in 2026 — 118 Findings Across 68 Packages
AI agents are installing packages on your machine every day. But how secure are those packages, really?
04
The State of MCP Server Security in 2026
118 findings across 194 audited packages — here's what we learned about the MCP ecosystem's security posture.
Use Cases
Four integration paths with different depth and intent — optimized for quick scanning and drill-down.
🛡️Fastest PathAgent Skill GuardrailPre-install gate
Inject package checks before install commands are executed by your assistant.
npx skills add agentaudit-dev/agentaudit-skill🔌In-Chat VerificationMCP Server Trust ChecksNatural language query
Ask for a trust score before your model connects to external MCP tooling.
"Is mcp-fetch safe?" → 97/100 PASS⚙️Pipeline PolicyCI/CD Security GateAutomated fail conditions
Stop merges when critical package findings appear in dependency diffs.
fail-on: critical📦Programmatic AccessRegistry & API LookupHTTP endpoint
Query package risk from scripts, bots, or internal security dashboards.
GET /api/check?package=expressAI Package Risk, Quantified
A trust-style dashboard showing exposure, critical findings, and current ecosystem coverage at a glance.
Getting Started
Install the AgentAudit Skill in one command. Works with all major AI platforms.
🥇 Recommended: Full Coverage
Auto-detect all IDEs (recommended)Auto-detects Claude Code, Cursor, Windsurf, Copilot, Codex, Kiro, and more
npx skills add agentaudit-dev/agentaudit-skillor install for a specific IDE
$npx skills add agentaudit-dev/agentaudit-skill --agent claude-code
Installs to
.claude/skills/ — add -g for global, -y to skip prompts✓ Pre-install checks✓ LLM code analysis✓ Submit findings
Full Docs →MCP Server
Ask your AI assistant "Is this package safe?" directly in chat.
$ git clone https://github.com/agentaudit-dev/agentaudit-mcp
$ npm install && npm run build
REST API
Look up any package directly. Free, no auth required for reads.
$ curl agentaudit.dev/api/check?package=express
Trust Score: 92/100 — PASS
GitHub Action
CI/CD security scanning. Flag or fail builds on unsafe packages.
uses: agentaudit-dev/agentaudit-github-action@v1
fail-on: critical
How AgentAudit keeps you informed
Three focused layers of security intelligence for every AI package in your stack.
01
Check Before Install
Search any package and get an instant trust score before code executes on your machine.
Trust ScoreStatic AnalysisDependency RiskInstall Verdict
@anthropic/mcp-fetch
97/100 — PASS ✅
02
Community Intelligence
New findings are pushed continuously and confidence improves through cross-validation.
- NEW ASF-2025-0142 · command injection
- UPDATE ASF-2025-0138 · score 72 → 45
- NEW ASF-2025-0144 · unsafe eval path
03
Verifiable Trust
Every score change and audit record is chain-linked, signed, and reproducible.
hash
a3f8c2...e91dprev
7b2e1a...f4c8sig
verified ✓Start securing your AI stack
One command. Every package checked before it runs.
npx skills add agentaudit-dev/agentaudit-skill