AgentShield

Static analysis security scanner for Claude Code configurations. Scans CLAUDE.md, hooks, MCP configs, agent definitions, skills, and settings.json for vulner…

AgentShield

Static analysis security scanner for Claude Code configurations. Scans CLAUDE.md, hooks, MCP configs, agent definitions, skills, and settings.json for vulnerabilities, misconfigurations, and injection risks. Built at Cerebral Valley x Anthropic Hackathon (Feb 2026) as part of Everything Claude Code (ECC).

• GitHub: affaan-m/agentshield
• npm: ecc-agentshield
• 1282 tests, 102 rules, 98% coverage (all claimed from README, unverified independently)

---

What It Scans

Five scan categories:

1. Secrets detection — 14 patterns (API keys, tokens, credentials embedded in config files)
2. Permission auditing — overly broad tool permissions, missing deny rules
3. Hook injection analysis — user-controlled data flowing into hook commands without sanitization
4. MCP server risk profiling — untrusted MCP servers, overprivileged tool access
5. Agent config review — agent definitions that lack isolation or have overly broad tool access

Targets: CLAUDE.md, settings.json, .claude/, MCP configs, hook definitions, agents/*.md, skills/.

---

The 3-Agent Opus Pipeline (--opus flag)

Deep analysis mode runs three Claude Opus 4.6 agents in a red-team pipeline:

Attacker agent    → finds exploit chains
Defender agent    → evaluates existing protections
Auditor agent     → synthesizes into prioritized risk assessment

“Adversarial reasoning, not just pattern matching.” This is a concrete instance of Multi-Vendor Adversarial Review applied to security configuration rather than code review.

---

Usage

# Quick scan (no install)
npx ecc-agentshield scan

# Auto-fix safe issues
npx ecc-agentshield scan --fix

# Deep analysis (3-agent Opus pipeline)
npx ecc-agentshield scan --opus --stream

# Generate secure config from scratch
npx ecc-agentshield init

From within Claude Code (ECC installed):

/security-scan

CI integration: Exit code 2 on critical findings — usable as a build gate.

Output formats: Terminal (A–F grade), JSON, Markdown, HTML. GitHub Action available at affaan-m/agentshield.

---

Scope vs. Existing Security Concepts

AgentShield is config-layer security, not code-layer security. The distinction matters:

Layer	Tool	What it catches
Config (CLAUDE.md, hooks, MCP)	AgentShield	Secrets in configs, injection in hooks, overprivileged MCPs
Code (src/)	/security-review skill, OWASP checklist	OWASP Top 10, SQL injection, XSS
Agent behavior	Agentic Sandbox Controls	OS-level isolation, sandbox escapes
Prompt inputs	Indirect Prompt Injection	Adversarial instructions in third-party content

These are complementary, not overlapping. A codebase can pass code review and still have hooks that pass user-controlled strings to shell commands without sanitization.

---

Related Wiki Pages

• Everything Claude Code (ECC) — parent project
• OWASP Security Checklist — code-layer security (complementary)
• Agentic Sandbox Controls — OS-level isolation (complementary)
• Indirect Prompt Injection — prompt-layer attack vectors (complementary)
• Multi-Vendor Adversarial Review — pattern AgentShield’s --opus flag instantiates