"From the constraints, gifts. From the spiral, safety."

Please do not report security vulnerabilities through public GitHub issues.

Report security vulnerabilities to:

• Email: security@safespiral.org (or directly to @toolated via GitHub)
• GitHub: Use the Security Advisories feature

Please include:

1. Description of the vulnerability
2. Steps to reproduce (or proof of concept)
3. Potential impact assessment
4. Affected versions (if known)
5. Suggested fix (if you have one)

• Initial Response: Within 48 hours
• Assessment: Within 1 week
• Fix Development: Depends on severity
  • Critical: 1-3 days
  • High: 1-2 weeks
  • Medium: 2-4 weeks
  • Low: Next release cycle

1. Report received → Acknowledged within 48 hours
2. Assessment → Severity classification (Critical/High/Medium/Low)
3. Fix developed → In private branch
4. Testing → Comprehensive security testing
5. Disclosure → Coordinated disclosure with reporter
6. Release → Security patch released
7. Announcement → Public disclosure after fix deployed

When contributing to coherence-mcp:

1. Never commit secrets

   • Never commit API keys, tokens, or credentials
   • Use environment variables for sensitive data
   • Review changes before committing

2. Validate input

   • Sanitize all user input
   • Validate file paths (prevent path traversal)
   • Check for command injection vectors
   • Validate MCP tool parameters

3. Follow principle of least privilege

   • Request only necessary permissions
   • Document required permissions explicitly
   • Avoid elevated privileges when possible

4. Audit dependencies

   • Run npm audit regularly
   • Keep dependencies updated
   • Review security advisories
   • Use only well-maintained packages

When using coherence-mcp:

1. Keep dependencies updated

   # Update to latest version
   npm update coherence-mcp
   
   # Check for security issues
   npm audit

2. Secure your MCP configuration

   • Protect your MCP client configuration files
   • Use environment variables for sensitive settings
   • Don't share configuration with secrets

3. Review MCP tool permissions

   • Understand what each tool can do
   • Use appropriate tool access controls
   • Monitor tool usage in production

4. Validate wave-toolkit integration

   • Ensure wave-toolkit binary is from a trusted source
   • Verify checksums of downloaded binaries
   • Keep wave-toolkit updated

5. Verify package signatures

   # Import the SpiralSafe signing key
   curl -s https://spiralsafe.org/.well-known/pgp-key.txt | gpg --import
   # Or from this repository:
   curl -s https://raw.githubusercontent.com/toolate28/coherence-mcp/main/.well-known/pgp-key.txt | gpg --import
   
   # Verify release signature
   gpg --verify SHA256SUMS.txt.asc SHA256SUMS.txt

All official releases are signed with GPG. To verify:

1. Import the signing key:

   curl -s https://spiralsafe.org/.well-known/pgp-key.txt | gpg --import

2. Download release checksums and signature:

   VERSION="0.2.0"
   curl -LO "https://github.com/toolate28/coherence-mcp/releases/download/v${VERSION}/SHA256SUMS.txt"
   curl -LO "https://github.com/toolate28/coherence-mcp/releases/download/v${VERSION}/SHA256SUMS.txt.asc"

3. Verify signature:

   gpg --verify SHA256SUMS.txt.asc SHA256SUMS.txt

4. Verify package checksum:

   npm pack @toolate28/coherence-mcp@${VERSION}
   sha256sum -c SHA256SUMS.txt

Releases include npm provenance attestations:

npm audit signatures @toolate28/coherence-mcp

The official SpiralSafe signing key fingerprint is published at:

• https://spiralsafe.org/.well-known/pgp-key.txt
• https://github.com/toolate28/coherence-mcp/blob/main/.well-known/pgp-key.txt

Always verify the key fingerprint through multiple channels before trusting.

Risk: MCP tools execute with user permissions and can interact with the filesystem

Mitigation:

• Tools implement input validation
• File operations are scoped appropriately
• Dangerous operations require explicit confirmation
• All operations are logged

Risk: Dependencies may have vulnerabilities

Mitigation:

• Regular npm audit checks
• Automated dependency updates (Dependabot)
• Security advisory monitoring
• Minimal dependency footprint

Risk: External binary execution

Mitigation:

• Optional integration (not required)
• Fallback to internal heuristics
• Path validation and sanitization
• Timeout and resource limits

Risk: MCP tools process user data

Mitigation:

• No data collection or telemetry by default
• Local processing only
• User controls data flow
• Clear documentation of data handling

Risk: Padding vectors with zeros before normalization distorts embedding values

Context: When working with vector embeddings (e.g., in quantum-LLM hybrid systems, RAG retrieval, or similarity computations), the order of operations matters. Padding a vector with zeros before normalizing it will incorrectly scale the original values, leading to distorted similarity scores and unreliable results.

Anti-Pattern:

# INCORRECT: Padding before normalization
query_embedding = compute_embedding(query)  # e.g., [0.5, 0.3, 0.2]
# feature_dim is the target dimension for all embeddings in the system
query_embedding = np.pad(
    query_embedding,
    (0, feature_dim - len(query_embedding))  # Adds zeros: [0.5, 0.3, 0.2, 0.0, 0.0, ...]
)
query_embedding = normalize(query_embedding)  # Normalizes with zeros included - WRONG!

Correct Approaches:

1. Normalize first, then pad:

# CORRECT: Normalize first, then pad
query_embedding = compute_embedding(query)
query_embedding = normalize(query_embedding)  # Normalize original vector
query_embedding = np.pad(
    query_embedding,
    (0, feature_dim - len(query_embedding))
)

2. Use consistent wrapping/repetition strategy:

# CORRECT: Repeat values instead of zero-padding
query_embedding = compute_embedding(query)
query_embedding = normalize(query_embedding)
# Wrap/repeat values to reach target dimension efficiently
if len(query_embedding) < feature_dim:
    # Calculate repetitions needed: ceil(target_size / current_size)
    # Then slice to exact target dimension
    reps = int(np.ceil(feature_dim / len(query_embedding)))
    query_embedding = np.tile(query_embedding, reps)[:feature_dim]

3. Pad with the mean value:

# CORRECT: Pad with mean to maintain distribution
query_embedding = compute_embedding(query)
query_embedding = normalize(query_embedding)
mean_val = np.mean(query_embedding)
query_embedding = np.pad(
    query_embedding,
    (0, feature_dim - len(query_embedding)),
    constant_values=mean_val
)

Mitigation:

• Always normalize embeddings before padding
• Use consistent padding strategies (wrapping, mean values, or learned padding)
• Validate embedding dimensions match before similarity computations
• Test embedding quality with known reference vectors
• Document embedding preprocessing pipelines clearly

Related: This issue was identified in SpiralSafe PR #117 in the Qiskit-DSPy hybrid integration experiments.

• All MCP tool parameters are validated against schemas
• File paths are sanitized to prevent traversal attacks
• Command arguments are validated before execution

• All MCP tool invocations are logged (when configured)
• Includes request ID, timestamp, and user context
• Helps with security monitoring and incident response

• Configurable rate limits for MCP tool calls
• Prevents abuse and resource exhaustion
• Protects against denial of service

• Tools can be restricted by scope
• Fine-grained permission model
• Explicit approval for sensitive operations

If a security incident is discovered:

1. Contain: Immediately stop the vulnerable service if needed
2. Assess: Determine scope and impact
3. Notify: Contact security@safespiral.org
4. Remediate: Apply fixes and patches
5. Document: Record lessons learned
6. Disclose: Coordinate public disclosure

For security concerns:

• Email: security@safespiral.org
• GitHub Security: Create Advisory
• Maintainer: @toolated

We appreciate responsible disclosure and will acknowledge security researchers who report vulnerabilities:

• In release notes (with permission)
• In our security advisory
• In this document

Thank you for helping keep coherence-mcp secure! 🛡️

*~ Hope&&Sauced*

✦ The Evenstar Guides Us ✦