A Look Inside Claude's Leaked AI Coding Agent

Key Takeaways

• Anthropic's Claude Code CLI source code was accidentally leaked via debug-only .map files in the public npm package @anthropic-ai/claude-code 2.1.88.
• The leak exposes the internal architecture, including the QueryEngine, tool execution flow, permission models, and system prompt assembly.
• Threat actors can analyze the exposed logic to bypass safety guardrails or craft targeted malicious repositories using prompt injection via CLAUDE.md files.
• Unofficial mirrors and reimplementations (e.g., instructkr/claw-code) have rapidly appeared, posing a significant risk of tampered dependencies for developers.

Affected Systems

• Claude Code CLI
• npm (@anthropic-ai/claude-code version 2.1.88)

Vulnerabilities (CVEs)

• Prompt Injection

Attack Chain

The incident originated from a human error where debug-only .map files were included in a public npm release of Claude Code. Developers and threat actors downloaded these files, extracting the full TypeScript source code. With access to the internal logic, attackers can analyze the permission models and system prompt assembly to craft malicious repositories containing prompt injections via CLAUDE.md files. Users downloading unofficial mirrors or tampered dependencies risk executing malicious code or bypassing AI safety guardrails on their local machines.

Detection Availability

• YARA Rules: No
• Sigma Rules: No
• Snort/Suricata Rules: No
• KQL Queries: No
• Splunk SPL Queries: No
• EQL Queries: No
• Other Detection Logic: No

The article does not provide specific detection rules or queries.

Detection Engineering Assessment

EDR Visibility: Low — EDR tools do not natively inspect LLM prompt injections or the internal logic of AI coding assistants unless they result in malicious child process execution. Network Visibility: Low — Traffic to Anthropic's API is encrypted and standard; malicious prompt injections would be embedded within legitimate API calls. Detection Difficulty: Hard — Detecting prompt injection within local CLAUDE.md files requires semantic analysis of the text, which standard signature-based tools cannot easily perform.

Required Log Sources

• npm audit logs
• Process execution logs
• File integrity monitoring (FIM)

Hunting Hypotheses

Control Gaps

• LLM Prompt Injection detection
• Supply chain verification for AI development tools

Key Behavioral Indicators

• Execution of Claude Code from non-official npm paths
• Unexpected modifications to CLAUDE.md containing suspicious system instructions

False Positive Assessment

• Low

Recommendations

Immediate Mitigation

• Ensure developers are only using the official @anthropic-ai/claude-code package.
• Audit internal projects for malicious CLAUDE.md files that could trigger prompt injections.

Infrastructure Hardening

• Implement strict egress filtering for development environments to prevent unauthorized code exfiltration or C2 communication from compromised AI agents.

User Protection

• Educate developers on the risks of downloading unofficial mirrors or reimplementations of leaked software.

Security Awareness

• Train engineering teams on AI prompt injection risks and secure usage of autonomous coding agents.

MITRE ATT&CK Mapping

• T1195.002 - Supply Chain Compromise: Compromise Software Supply Chain
• T1562.001 - Impair Defenses: Disable or Modify Tools
• T1588 - Obtain Capabilities

Additional IOCs

• File Paths:
  • QueryEngine.ts - Core engine file exposed in the leak handling LLM interaction lifecycle.
  • Tool.ts - Core file exposed in the leak handling tool execution.
  • commands.ts - Core file exposed in the leak handling command systems.
  • CLAUDE.md - Project-level context file used by Claude Code, identified as a potential vector for prompt injection attacks.