MIND is designed with specific security properties:

MIND assumes:

• The host process is trusted
• Embeddings come from trusted sources
• State files are stored securely
• The runtime environment is not compromised

MIND does not defend against:

• Malicious embeddings designed to poison memory
• Compromised host processes
• Physical access attacks
• Memory inspection by other processes

Do not report security vulnerabilities through public GitHub issues.

Instead, please report them via email to:

security@mind-project.org

Include:

1. Description of the vulnerability
2. Steps to reproduce
3. Potential impact
4. Suggested fix (if any)

1. Acknowledgment — We will confirm receipt of your report
2. Assessment — We will evaluate severity and impact
3. Communication — We will keep you informed of progress
4. Credit — We will credit you in the advisory (unless you prefer anonymity)

When using MIND:

# Restrict permissions
chmod 600 mind.state

# Store in secure location
mv mind.state /secure/path/

# Consider encryption at rest

• Validate embeddings before feeding to MIND
• Use trusted embedding models
• Monitor for anomalous patterns

• Run MIND in isolated processes
• Use sandboxing where available
• Limit file system access
• Monitor resource usage

MIND welcomes security audits. If you are conducting a security assessment:

1. Contact us at security@mind-project.org
2. We will provide a point of contact
3. We can arrange access to development builds
4. We will coordinate disclosure

When all memory slots are full, new experiences are silently ignored. This is by design (bounded memory) but could be exploited to prevent learning.

Mitigation: Monitor slot utilization via cr_state_temporal().

Plasticity never reaches zero (ε = 0.05), meaning the system is always slightly malleable. This is by design (mercy) but means beliefs can always be influenced given enough repetition.

Mitigation: This is a feature, not a bug. Design your system accordingly.

S2S calibration signals are not authenticated. A malicious peer could send false calibration data.

Mitigation: Authenticate peers at the transport layer. MIND does not handle authentication.