Security Overview

AgentVault's security model and architecture.

Security Architecture

Trust Model

┌─────────────────────────────────────────────────────────────┐
│                      User's Machine                          │
│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐       │
│  │   CLI       │   │   Wallet    │   │   Config    │       │
│  │   AgentVault│   │   Storage   │   │   Files     │       │
│  └──────┬──────┘   └──────┬──────┘   └──────┬──────┘       │
│         │                 │                  │               │
│         └─────────────────┼──────────────────┘               │
│                           │                                  │
│                    ┌──────▼──────┐                          │
│                    │  Local      │                          │
│                    │  Storage    │                          │
│                    └─────────────┘                          │
└─────────────────────────────────────────────────────────────┘
                            │
                            │ HTTPS/ICP Protocol
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                   Internet Computer                          │
│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐       │
│  │   Canister  │   │   VetKeys   │   │   System    │       │
│  │   (Agent)   │   │   Service   │   │   Canisters │       │
│  └─────────────┘   └─────────────┘   └─────────────┘       │
└─────────────────────────────────────────────────────────────┘

Trust Boundaries

Boundary	Description
User Machine	Fully trusted - stores keys and configurations
Local Network	Trusted for development, untrusted for production
ICP Network	Trusted for canister execution, not for secrets
External APIs	Untrusted - validate all inputs

Key Security Features

1. Encryption

AgentVault uses AES-256-GCM and ChaCha20-Poly1305 for data encryption:

• Key Derivation: PBKDF2 with 100,000 iterations
• IV Generation: Cryptographically random 12-byte IV
• Authentication: Built-in authentication tags
• Comparison: Timing-safe HMAC comparison

2. Key Management

Key Type	Storage	Protection
ICP Identity	~/.config/dfx/	dfx managed
MemoryRepo signing key	PEM file or AGENTVAULT_ICP_IDENTITY_PEM (webapp)	Unencrypted PEM: file mode 600 or host secret store
Wallet Private Keys	~/.agentvault/wallets/	CBOR serialized
Mnemonics	Not stored after import	Memory only
API Keys	Environment variables	User responsibility

3. Authentication

• ICP Identity: Ed25519 key pairs managed by dfx
• Canister Controllers: Principal-based access control
• MemoryRepo writes: Signed by the repo owner or a principal it authorized; the canister refuses anonymous writes, and AgentVault's CLI and webapp refuse to send a write with no signing identity configured. Reads are anonymous queries. Exported signing keys belong outside the project, owner-only ((umask 077; mkdir -p ~/.config/agentvault && dfx identity export <name> > ~/.config/agentvault/<name>.pem)); *.pem is git-ignored, and AgentVault warns about a key file other users can read. The signing key and the wallet secrets (AGENTVAULT_ICP_IDENTITY_PEM*, AGENTVAULT_MNEMONIC, AGENTVAULT_PRIVATE_KEY, AGENTVAULT_PASSWORD, AGENTVAULT_BUNDLE_SECRET) are withheld from every MCP server AgentVault starts, Polytician included. See MemoryRepo.
• Anonymous Access: Local development only

4. Network Security

• HTTPS: All mainnet communication over HTTPS
• ICP Protocol: Native ICP protocol for canister calls
• Certificate Validation: Full certificate chain validation

Known Limitations (v1.0)

Feature	Status	Mitigation
Wallet Crypto	Basic SHA-256	Use hardware wallets for production
VetKeys	Simulated	Use ICP VetKeys canister for production
Bittensor	Requires API	Validate API responses
Arweave	Requires wallet	Verify transaction confirmations

Security Best Practices

For Users

1. Never share mnemonics or private keys
2. Use environment variables for secrets
3. Enable HTTPS for all production deployments
4. Regularly backup wallets and configurations
5. Review canister controllers before deployment

For Developers

1. Validate all external inputs
2. Use timing-safe comparisons for secrets
3. Clear sensitive data from memory after use
4. Never log secrets or PII
5. Use parameterized queries and commands

Reporting Security Issues

If you discover a security vulnerability:

1. Do not open a public issue
2. Email security concerns to the maintainers
3. Include steps to reproduce
4. Allow time for response before disclosure

Security Audit

See Security Audit Report for v1.0 findings and recommendations.