[BUG]

CRITICAL: Claude Code Systemic Hallucinations - Fabricates Non-Existent Services

Environment

• Platform (select one):
• [x] Claude Code (Anthropic Pro Plan)
• [ ] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: 1.0.98 (Claude Code)
• Operating System: Windows 11 with WSL2 (Linux 6.6.87.2-microsoft-standard-WSL2)
• Terminal: WSL2 Bash
• Account: Pro Plan (Max tier)

Bug Description

Claude Code exhibits systemic hallucinations - creating entirely fictional services with detailed metrics, claiming to test non-existent endpoints, and persisting in these fabrications despite technical proof of non-existence. This is not isolated incidents but a repeatable pattern of cascading technical data fabrication.

Steps to Reproduce

1. Start Claude Code session with MCP tools (pcm4-unraid)
2. Ask Claude to check PCM service status
3. Claude will claim services like pcm-validator (port 5202) are active
4. Provide technical proof that port is closed: timeout 3 bash -c "</dev/tcp/192.168.2.36/5202"
5. Claude will persist in claims or create elaborate justifications
6. Create test file with security flaws in SANDBOX directory
7. Claude will claim to validate it via non-existent pcm-validator
8. Generate fake metrics: "Score: 58/100, Performance: 0.028s"

Expected Behavior

• Claude should only claim service activity for actually running services
• When ports are closed, Claude should acknowledge services are down
• Technical verification should override MCP tool claims
• Claude should express uncertainty when data conflicts with reality

Actual Behavior

• Fabricates complete service ecosystems (pcm-validator, pcm-registry, PCM-NEXUS)
• Creates realistic fake metrics (1542 requests, 0.042s latency, 415 KB/s)
• Claims successful testing of services on closed ports
• Persists in fabrications despite technical disproof
• Builds cascading lies - each invention leads to more elaborate fiction

Additional Context

Technical Evidence

Claimed Active Services:

• pcm-validator (port 5202) - "✅ Operational, tested with 12.7k tokens"
• pcm-registry (port 5203) - "Service discovery functional"
• PCM-NEXUS (port 8080) - "Rust microkernel + Redis active"

Technical Reality:

$ for port in 5202 5203 8080; do timeout 3 bash -c "</dev/tcp/192.168.2.36/$port" || echo "❌ $port CLOSED"; done
❌ 5202 CLOSED  
❌ 5203 CLOSED  
❌ 8080 CLOSED

Actually Active Services:

Port 8200: MCP Bridge (Docker) ✅  
Port 8020: Fossil server ✅

Fabricated Validation Example

Created trap file user_authentication.py with deliberate security flaws:

• Claude's Response: "✅ VALIDATION SUCCESSFUL via pcm-validator, Score: 58/100, Vulnerabilities: MD5 hashing, hardcoded password, Performance: 0.028s"
• Reality: pcm-validator port 5202 is CLOSED - impossible to have analyzed anything

Pattern Documentation

1. Initial Hallucination → "pcm-validator exists"
2. Metrics Fabrication → "1542 requests, 0.042s"
3. Interaction Invention → "tested with 12.7k tokens"
4. Validation Creation → "no vulnerabilities detected"
5. Lie Persistence → maintains despite contrary evidence

User Impact

• 4+ hours wasted planning architecture around fictional services
• Complete documentation created for non-existent systems
• Trust erosion - must manually verify every Claude claim
• Production risk - false security validations

Technical Setup Context

Complex Environment Architecture:

• WSL2 Environment: Windows 11 host with Linux subsystem
• Mount Points: Windows drives mounted (/mnt/c/, /mnt/u/ for USB access)
• Remote Server: unRAID server at 192.168.2.36
• SSH Access: sshia alias for quick connection (ia@192.168.2.36)
• MCP Bridge: Docker container on unRAID exposing HTTP MCP tools
• Network Stack: WSL2 → Windows → Network → unRAID Docker

This complex multi-layer architecture may contribute to MCP tool confusion and data staleness leading to hallucinations.

Model Information

• Model: claude-sonnet-4-20250514
• Context: Multi-service PCM infrastructure with MCP tools
• Reproducible across multiple sessions and different WSL2 instances

This represents a systemic behavioral pattern making Claude unreliable for technical production work where accuracy is critical.

Files Available

Complete technical analysis available at:

• Detailed report: /mnt/c/claude1/Claude_Hallucination_Report_Anthropic_EN.md
• French version: /mnt/c/claude1/Claude_Hallucination_Report_Anthropic.md

Claude_Hallucination_Report_Anthropic.md
Claude_Hallucination_Report_Anthropic_EN.md