DO NOT create public GitHub issues for security vulnerabilities.

• Email: security@solana-immune.system
• PGP Key: [Available on request]

We offer bounties for responsible disclosure:

In Scope:

• Smart contracts (programs/)
• Agent logic (agents/)
• API endpoints (app/api/)

Out of Scope:

• Third-party dependencies
• Documentation issues
• Social engineering

┌─────────────────────────────────────────────────────────────┐
│                    SECURITY LAYERS                          │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Layer 1: INPUT VALIDATION                                  │
│  ├── All inputs sanitized                                   │
│  ├── Rate limiting on all endpoints                         │
│  └── Schema validation                                      │
│                                                             │
│  Layer 2: HUMAN OVERSIGHT                                   │
│  ├── High-severity actions require human approval           │
│  ├── Emergency stop mechanisms                              │
│  └── Audit trail for all decisions                          │
│                                                             │
│  Layer 3: CONSENSUS REQUIREMENTS                            │
│  ├── Multi-agent voting for critical actions                │
│  ├── Minimum 3 agents must agree                            │
│  └── Coordinator can veto dangerous actions                 │
│                                                             │
│  Layer 4: ON-CHAIN VERIFICATION                             │
│  ├── All reasoning committed before action                  │
│  ├── Hash verification prevents tampering                   │
│  └── Public audit trail on Solana                           │
│                                                             │
│  Layer 5: FALLBACK SYSTEMS                                  │
│  ├── Deterministic rules if AI unavailable                  │
│  ├── Graceful degradation                                   │
│  └── Manual override always available                       │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Critical Actions Requiring Human Approval:

• Blocking addresses with >$10,000 in assets
• Coordinated responses involving >5 agents
• Actions affecting >100 users
• First-time threat patterns (no historical match)

Emergency Controls:

• EMERGENCY_STOP environment variable halts all agents
• Admin can pause individual agents via API
• All actions can be reversed within 1 hour

1. Deterministic Fallbacks: If Claude API unavailable, system uses rule-based detection
2. Confidence Thresholds: Actions only taken if confidence >70%
3. False Positive Protection: Requires 3+ agent agreement for blocking
4. Rate Limiting: Max 10 blocking actions per hour
5. Cooldown Periods: Same address cannot be re-evaluated for 5 minutes

• No upgradeable proxies: Contracts are immutable after deployment
• Minimal permissions: Each contract has single responsibility
• PDA-based authority: No admin keys that can be compromised
• Tested on devnet: Extensive testing before mainnet

• Internal Review: Complete
• Automated Scanning: Passing (Clippy, cargo-audit)
• Test Coverage: See coverage reports

• Professional audit by OtterSec (Q2 2026)
• Formal verification of critical paths
• Penetration testing

1. AI Dependency: System degrades (but doesn't fail) without Claude API
2. Centralized RPC: Currently uses Helius (planning to add fallback RPCs)
3. New Project: Limited production history

1. Detection: Automated monitoring + community reports
2. Triage: Severity assessment within 1 hour
3. Mitigation: Emergency stop if needed
4. Resolution: Fix deployed within 24 hours for critical issues
5. Disclosure: Public post-mortem after resolution

• Open source (MIT License)
• No user funds custody
• Transparent decision-making
• GDPR-compliant logging