BACK_TO_ARCHIVE

7 Vulnerabilities in Claude Code's Internal Architecture

Static Analysis of a Leaked AI Assistant CLI — Permission Bypass, SSRF, Prompt Injection, and More

Advanced22 min read2026-03

Vulnerabilities Found

7

Attack Layers

4

Time to Report

24h

01_THE_CHALLENGE
The leaked source (cli.js.map) exposed the full TypeScript architecture of an AI CLI handling arbitrary file system access, shell execution, and external HTTP requests — all gated behind a permission system. The challenge: systematically audit every trust boundary before the window closed.
02_THE_SOLUTION
Methodical layer-by-layer static analysis: • Prompt injection defense — CLAUDE.md include chain, nested file loading • Permission system — TOCTOU races, symlink resolution, classifier gaps • Tool execution — WebFetch SSRF vectors, Skill tool argument injection • Infrastructure — upstream proxy CA certificate trust • Information disclosure — undercover mode sanitization gaps Each finding includes a reproduction path, CVSS-style impact assessment, and a concrete fix proposal.
03_IMPACT_METRICS

Technical_Impact

  • 7 vulnerabilities identified across 4 distinct attack layers
  • CC-001: CLAUDE.md nested includes enable prompt injection via repo content
  • CC-002: Symlink resolution TOCTOU — permission check and file read are non-atomic
  • CC-003: Upstream proxy CA injection — attacker-controlled cert trusted at CCR layer
  • CC-004: Bash classifier absent in external builds — permission bypass in non-Anthropic forks
  • CC-005: WebFetch SSRF via crafted query parameters reaching internal endpoints
  • CC-006: Skill tool argument injection — user-supplied args reach prompt without sanitization
  • CC-007: Undercover mode leaks partial identity through response phrasing patterns

Business_Impact

  • Responsible disclosure filed — no public exploit released
  • Research published on Al Bayrouni before any competing analysis
  • Demonstrates AI system security auditing as an emerging discipline
  • Methodology reusable across any AI CLI with tool execution capabilities
04_TECHNICAL_DEEP_DIVE
05_TECH_STACK
Security ResearchStatic AnalysisPrompt InjectionSSRFPermission SystemsAI SecurityResponsible Disclosure

Want the full technical breakdown?

The wire-format analysis, architecture diagrams, and protocol-level detail live on Al Bayrouni. The contact form is for consulting and engagement discussions.

Rahmouni Oussama | Senior Full-Stack Engineer & Architect