This document provides a comprehensive technical reference for every security system in the OpenFang Agent Operating System. All struct names, function signatures, constant values, and algorithm descriptions are drawn directly from the source code.
- Security Overview
- Capability-Based Security
- WASM Dual Metering
- Merkle Hash Chain Audit Trail
- Information Flow Taint Tracking
- Ed25519 Manifest Signing
- SSRF Protection
- Secret Zeroization
- OFP Mutual Authentication
- Security Headers
- GCRA Rate Limiter
- Path Traversal Prevention
- Subprocess Sandbox
- Prompt Injection Scanner
- Loop Guard
- Session Repair
- Health Endpoint Redaction
- Security Configuration
- Security Dependencies
OpenFang implements defense-in-depth security. No single mechanism is trusted to be the sole protector; instead, 16 independent systems form overlapping layers so that a failure in any one layer is caught by others.
| # | System | Crate | Protects Against |
|---|---|---|---|
| 1 | Capability-Based Security | openfang-types |
Unauthorized actions by agents |
| 2 | WASM Dual Metering | openfang-runtime |
Infinite loops, CPU DoS |
| 3 | Merkle Audit Trail | openfang-runtime |
Tampered audit logs |
| 4 | Taint Tracking | openfang-types |
Prompt injection, data exfiltration |
| 5 | Ed25519 Manifest Signing | openfang-types |
Supply chain attacks |
| 6 | SSRF Protection | openfang-runtime |
Server-Side Request Forgery |
| 7 | Secret Zeroization | openfang-runtime, openfang-channels |
Memory forensics, key leakage |
| 8 | OFP Mutual Auth | openfang-wire |
Unauthorized peer connections |
| 9 | Security Headers | openfang-api |
XSS, clickjacking, MIME sniffing |
| 10 | GCRA Rate Limiter | openfang-api |
API abuse, denial of service |
| 11 | Path Traversal Prevention | openfang-runtime |
Directory traversal attacks |
| 12 | Subprocess Sandbox | openfang-runtime |
Secret leakage via child processes |
| 13 | Prompt Injection Scanner | openfang-skills |
Malicious skill prompts |
| 14 | Loop Guard | openfang-runtime |
Stuck agent tool loops |
| 15 | Session Repair | openfang-runtime |
Corrupted LLM conversation history |
| 16 | Health Endpoint Redaction | openfang-api |
Information leakage |
Source: openfang-types/src/capability.rs
OpenFang uses capability-based security. An agent can only perform actions it has been explicitly granted permission to do. Capabilities are immutable after agent creation and are enforced at the kernel level.
The Capability enum defines every permission type:
pub enum Capability {
// Filesystem
FileRead(String), // Glob pattern, e.g. "/data/*"
FileWrite(String),
// Network
NetConnect(String), // Host:port pattern, e.g. "*.openai.com:443"
NetListen(u16),
// Tools
ToolInvoke(String), // Specific tool ID
ToolAll, // All tools (dangerous)
// LLM
LlmQuery(String),
LlmMaxTokens(u64),
// Agent interaction
AgentSpawn,
AgentMessage(String),
AgentKill(String),
// Memory
MemoryRead(String),
MemoryWrite(String),
// Shell
ShellExec(String),
EnvRead(String),
// OFP Wire Protocol
OfpDiscover,
OfpConnect(String),
OfpAdvertise,
// Economic
EconSpend(f64),
EconEarn,
EconTransfer(String),
}The capability_matches(granted, required) function implements glob-style
matching:
- Exact match:
"api.openai.com:443"matches"api.openai.com:443" - Full wildcard:
"*"matches anything - Prefix wildcard:
"*.openai.com:443"matches"api.openai.com:443" - Suffix wildcard:
"api.*"matches"api.openai.com" - Middle wildcard:
"api.*.com"matches"api.openai.com" - ToolAll special case:
ToolAllgrants anyToolInvoke(_) - Numeric bounds:
LlmMaxTokens(10000)grantsLlmMaxTokens(5000)(granted >= required)
In the WASM sandbox, every host call is checked before execution by
check_capability() in host_functions.rs:
fn check_capability(
capabilities: &[Capability],
required: &Capability,
) -> Result<(), serde_json::Value> {
for granted in capabilities {
if capability_matches(granted, required) {
return Ok(());
}
}
Err(json!({"error": format!("Capability denied: {required:?}")}))
}If no granted capability matches the required one, the operation returns a JSON error immediately -- the tool is never invoked.
When an agent spawns a child agent, validate_capability_inheritance() ensures
the child's capabilities are a subset of the parent's. This prevents
privilege escalation:
pub fn validate_capability_inheritance(
parent_caps: &[Capability],
child_caps: &[Capability],
) -> Result<(), String> {
for child_cap in child_caps {
let is_covered = parent_caps
.iter()
.any(|parent_cap| capability_matches(parent_cap, child_cap));
if !is_covered {
return Err(format!(
"Privilege escalation denied: child requests {:?} \
but parent does not have a matching grant",
child_cap
));
}
}
Ok(())
}The host_agent_spawn() function in host_functions.rs calls
kernel.spawn_agent_checked(manifest_toml, Some(&state.agent_id), &state.capabilities)
which invokes this validation before the child is created.
Source: openfang-runtime/src/sandbox.rs
Untrusted WASM modules run inside a Wasmtime sandbox with two independent metering mechanisms running simultaneously.
Fuel metering counts WASM instructions. The engine deducts fuel for every
instruction executed. When the budget is exhausted, execution traps with
Trap::OutOfFuel.
// SandboxConfig defaults
pub fuel_limit: u64, // Default: 1_000_000
// Applied at execution time
if config.fuel_limit > 0 {
store.set_fuel(config.fuel_limit)?;
}After execution, fuel consumed is reported:
let fuel_remaining = store.get_fuel().unwrap_or(0);
let fuel_consumed = config.fuel_limit.saturating_sub(fuel_remaining);A watchdog thread sleeps for the configured timeout, then increments the
engine epoch. When the epoch advances past the store's deadline, execution
traps with Trap::Interrupt.
store.set_epoch_deadline(1);
let engine_clone = engine.clone();
let timeout = config.timeout_secs.unwrap_or(30);
let _watchdog = std::thread::spawn(move || {
std::thread::sleep(std::time::Duration::from_secs(timeout));
engine_clone.increment_epoch();
});| Property | Fuel | Epoch |
|---|---|---|
| Metric | Instruction count | Wall-clock time |
| Precision | Deterministic, reproducible | Non-deterministic |
| Catches | CPU-intensive loops | Host call blocking, I/O waits |
| Evasion | Can waste time in host calls | Can busy-loop cheaply |
Together they form a complete defense: fuel catches compute-intensive loops, while epochs catch host-call abuse or environmental slowdowns.
pub struct SandboxConfig {
pub fuel_limit: u64, // Default: 1_000_000
pub max_memory_bytes: usize, // Default: 16 MB
pub capabilities: Vec<Capability>,
pub timeout_secs: Option<u64>, // Default: 30 seconds
}pub enum SandboxError {
Compilation(String),
Instantiation(String),
Execution(String),
FuelExhausted, // Trap::OutOfFuel
AbiError(String),
}Source: openfang-runtime/src/audit.rs
Every security-critical action is appended to a tamper-evident Merkle hash chain, similar to a blockchain. Each entry contains the SHA-256 hash of its own contents concatenated with the hash of the previous entry.
pub enum AuditAction {
ToolInvoke,
CapabilityCheck,
AgentSpawn,
AgentKill,
AgentMessage,
MemoryAccess,
FileAccess,
NetworkAccess,
ShellExec,
AuthAttempt,
WireConnect,
ConfigChange,
}pub struct AuditEntry {
pub seq: u64, // Monotonically increasing sequence number
pub timestamp: String, // ISO-8601
pub agent_id: String,
pub action: AuditAction,
pub detail: String, // e.g. tool name, file path
pub outcome: String, // "ok", "denied", error message
pub prev_hash: String, // SHA-256 of previous entry (or 64 zeros)
pub hash: String, // SHA-256 of this entry + prev_hash
}Each entry's hash is computed from all of its fields concatenated with the previous entry's hash:
fn compute_entry_hash(
seq: u64, timestamp: &str, agent_id: &str,
action: &AuditAction, detail: &str,
outcome: &str, prev_hash: &str,
) -> String {
let mut hasher = Sha256::new();
hasher.update(seq.to_string().as_bytes());
hasher.update(timestamp.as_bytes());
hasher.update(agent_id.as_bytes());
hasher.update(action.to_string().as_bytes());
hasher.update(detail.as_bytes());
hasher.update(outcome.as_bytes());
hasher.update(prev_hash.as_bytes());
hex::encode(hasher.finalize())
}AuditLog::verify_integrity() walks the entire chain and recomputes every
hash. If any entry has been tampered with, the recomputed hash will not match
the stored hash, or the prev_hash linkage will be broken:
pub fn verify_integrity(&self) -> Result<(), String> {
let entries = self.entries.lock().unwrap_or_else(|e| e.into_inner());
let mut expected_prev = "0".repeat(64); // Genesis sentinel
for entry in entries.iter() {
if entry.prev_hash != expected_prev {
return Err(format!(
"chain break at seq {}: expected prev_hash {} but found {}",
entry.seq, expected_prev, entry.prev_hash
));
}
let recomputed = compute_entry_hash(/* ... */);
if recomputed != entry.hash {
return Err(format!(
"hash mismatch at seq {}: expected {} but found {}",
entry.seq, recomputed, entry.hash
));
}
expected_prev = entry.hash.clone();
}
Ok(())
}AuditLog uses Mutex<Vec<AuditEntry>> and Mutex<String> for the tip hash.
Both locks use unwrap_or_else(|e| e.into_inner()) to recover from poisoned
mutexes, ensuring the audit log remains available even after a panic.
| Method | Description |
|---|---|
AuditLog::new() |
Creates an empty log with genesis sentinel ("0" * 64) |
record(agent_id, action, detail, outcome) |
Appends an entry, returns its hash |
verify_integrity() |
Validates the entire chain |
tip_hash() |
Returns the hash of the most recent entry |
len() / is_empty() |
Entry count |
recent(n) |
Returns the most recent n entries (cloned) |
Source: openfang-types/src/taint.rs
OpenFang implements a lattice-based taint propagation model that prevents tainted values from flowing into sensitive sinks without explicit declassification. This guards against prompt injection, data exfiltration, and confused-deputy attacks.
pub enum TaintLabel {
ExternalNetwork, // Data from external network requests
UserInput, // Direct user input
Pii, // Personally identifiable information
Secret, // API keys, tokens, passwords
UntrustedAgent, // Data from sandboxed/untrusted agents
}pub struct TaintedValue {
pub value: String, // The payload
pub labels: HashSet<TaintLabel>, // Attached taint labels
pub source: String, // Human-readable origin
}Key methods:
| Method | Description |
|---|---|
TaintedValue::new(value, labels, source) |
Create with labels |
TaintedValue::clean(value, source) |
Create with no labels (untainted) |
merge_taint(&mut self, other) |
Union of labels (for concatenation) |
check_sink(&self, sink) |
Check if value can flow to sink |
declassify(&mut self, label) |
Remove a specific label (explicit security decision) |
is_tainted(&self) -> bool |
True if any labels present |
A TaintSink defines which labels are blocked from reaching it:
| Sink | Blocked Labels | Rationale |
|---|---|---|
TaintSink::shell_exec() |
ExternalNetwork, UntrustedAgent, UserInput |
Prevents command injection |
TaintSink::net_fetch() |
Secret, Pii |
Prevents data exfiltration |
TaintSink::agent_message() |
Secret |
Prevents secret leakage to other agents |
When check_sink() finds a blocked label, it returns a TaintViolation:
pub struct TaintViolation {
pub label: TaintLabel, // The offending label
pub sink_name: String, // "shell_exec", "net_fetch", etc.
pub source: String, // Where the tainted value came from
}Display: taint violation: label 'Secret' from source 'env_var' is not allowed to reach sink 'net_fetch'
Declassification is an explicit security decision. The caller asserts that the value has been sanitized:
tainted.declassify(&TaintLabel::ExternalNetwork);
tainted.declassify(&TaintLabel::UserInput);
// After declassification, value can flow to shell_exec
assert!(tainted.check_sink(&TaintSink::shell_exec()).is_ok());When two values are combined (concatenation, interpolation), the result must carry the union of both label sets:
let mut combined = TaintedValue::new(/* ... */);
combined.merge_taint(&other_value);
// combined.labels is now the union of bothSource: openfang-types/src/manifest_signing.rs
Agent manifests define an agent's capabilities, tools, and configuration. A compromised manifest can grant elevated privileges. This module provides Ed25519-based cryptographic signing.
- Compute SHA-256 of the manifest content (raw TOML text).
- Sign the hash with Ed25519 (via
ed25519-dalek). - Bundle the signature, public key, and content hash into a
SignedManifestenvelope.
pub struct SignedManifest {
pub manifest: String, // Raw TOML content
pub content_hash: String, // Hex SHA-256 of manifest
pub signature: Vec<u8>, // Ed25519 signature (64 bytes)
pub signer_public_key: Vec<u8>, // Ed25519 public key (32 bytes)
pub signer_id: String, // Human-readable signer ID
}let signing_key = SigningKey::generate(&mut OsRng);
let signed = SignedManifest::sign(manifest_toml, &signing_key, "admin@org.com");Internally:
pub fn sign(manifest: impl Into<String>, signing_key: &SigningKey, signer_id: impl Into<String>) -> Self {
let manifest = manifest.into();
let content_hash = hash_manifest(&manifest); // SHA-256
let signature = signing_key.sign(content_hash.as_bytes());
let verifying_key = signing_key.verifying_key();
Self {
manifest,
content_hash,
signature: signature.to_bytes().to_vec(),
signer_public_key: verifying_key.to_bytes().to_vec(),
signer_id: signer_id.into(),
}
}Two-phase verification:
- Hash check: Recompute SHA-256 of
manifestand compare tocontent_hash. - Signature check: Verify the Ed25519 signature over
content_hashusingsigner_public_key.
pub fn verify(&self) -> Result<(), String> {
let recomputed = hash_manifest(&self.manifest);
if recomputed != self.content_hash {
return Err("content hash mismatch: ...");
}
let verifying_key = VerifyingKey::from_bytes(&pk_bytes)?;
let signature = Signature::from_bytes(&sig_bytes);
verifying_key.verify(self.content_hash.as_bytes(), &signature)
.map_err(|e| format!("signature verification failed: {}", e))
}- Modifying the manifest content after signing causes a content hash mismatch.
- Replacing the public key with a different key causes a signature verification failure.
- Both attacks are caught by
verify().
Source: openfang-runtime/src/host_functions.rs
The host_net_fetch function (WASM host call for network requests) includes
comprehensive Server-Side Request Forgery protection.
Only http:// and https:// schemes are allowed. All others (file://,
gopher://, ftp://) are blocked immediately:
if !url.starts_with("http://") && !url.starts_with("https://") {
return Err(json!({"error": "Only http:// and https:// URLs are allowed"}));
}Before DNS resolution, these hostnames are blocked:
localhostmetadata.google.internalmetadata.aws.internalinstance-data169.254.169.254(AWS/GCP metadata endpoint)
After the hostname blocklist, the function resolves the hostname to IP addresses and checks every resolved IP against private ranges. This defeats DNS rebinding attacks:
let socket_addr = format!("{hostname}:{port}");
if let Ok(addrs) = socket_addr.to_socket_addrs() {
for addr in addrs {
let ip = addr.ip();
if ip.is_loopback() || ip.is_unspecified() || is_private_ip(&ip) {
return Err(json!({"error": format!(
"SSRF blocked: {hostname} resolves to private IP {ip}"
)}));
}
}
}The is_private_ip() function covers:
IPv4:
10.0.0.0/8-- RFC 1918172.16.0.0/12-- RFC 1918192.168.0.0/16-- RFC 1918169.254.0.0/16-- Link-local (AWS metadata)
IPv6:
fc00::/7-- Unique Local Addressfe80::/10-- Link-local
fn is_private_ip(ip: &std::net::IpAddr) -> bool {
match ip {
IpAddr::V4(v4) => {
let octets = v4.octets();
matches!(
octets,
[10, ..] | [172, 16..=31, ..] | [192, 168, ..] | [169, 254, ..]
)
}
IpAddr::V6(v6) => {
let segments = v6.segments();
(segments[0] & 0xfe00) == 0xfc00 || (segments[0] & 0xffc0) == 0xfe80
}
}
}extract_host_from_url() parses the URL to extract host:port for both
SSRF checking and capability matching:
https://api.openai.com/v1/chat -> api.openai.com:443
http://localhost:8080/api -> localhost:8080
http://example.com -> example.com:80
Source: All LLM driver modules, channel adapters, and web search modules.
OpenFang uses Zeroizing<String> from the zeroize crate on every field
that holds secret material. When the value is dropped, its memory is
overwritten with zeros, preventing secrets from lingering in memory.
Zeroizing<T> is a smart-pointer wrapper from the zeroize crate. It
implements Deref<Target=T> for transparent usage and Drop for automatic
zeroization:
// On Drop, the inner String's buffer is overwritten with zeros
let key = Zeroizing::new("sk-secret-key".to_string());
// Use key transparently via Deref
client.post(url).header("authorization", format!("Bearer {}", &*key));
// When key goes out of scope, memory is zeroedLLM Drivers (openfang-runtime/src/drivers/):
| Driver | Field |
|---|---|
AnthropicDriver |
api_key: Zeroizing<String> |
GeminiDriver |
api_key: Zeroizing<String> |
OpenAiCompatDriver |
api_key: Zeroizing<String> |
Channel Adapters (openfang-channels/src/):
| Adapter | Field(s) |
|---|---|
DiscordAdapter |
token: Zeroizing<String> |
EmailAdapter |
password: Zeroizing<String> |
BlueskyAdapter |
app_password: Zeroizing<String> |
DingTalkAdapter |
access_token: Zeroizing<String>, secret: Zeroizing<String> |
FeishuAdapter |
app_secret: Zeroizing<String> |
FlockAdapter |
bot_token: Zeroizing<String> |
GitterAdapter |
token: Zeroizing<String> |
GotifyAdapter |
app_token: Zeroizing<String>, client_token: Zeroizing<String> |
Web Search (openfang-runtime/src/web_search.rs):
fn resolve_api_key(env_var: &str) -> Option<Zeroizing<String>> {
std::env::var(env_var).ok().filter(|k| !k.is_empty()).map(Zeroizing::new)
}Embedding (openfang-runtime/src/embedding.rs):
| Struct | Field |
|---|---|
EmbeddingClient |
api_key: Zeroizing<String> |
Without zeroization, secrets remain in memory after use until the OS
reclaims the page. An attacker with access to a core dump, swap file, or
memory forensics tool can recover API keys. Zeroizing<String> ensures
the secret is overwritten as soon as it is no longer needed.
Source: openfang-wire/src/peer.rs
The OpenFang Wire Protocol (OFP) uses HMAC-SHA256 with nonce-based mutual authentication over TCP connections.
OFP refuses to start without a shared_secret:
if config.shared_secret.is_empty() {
return Err(WireError::HandshakeFailed(
"OFP requires shared_secret. Set [network] shared_secret in config.toml".into(),
));
}type HmacSha256 = Hmac<Sha256>;
fn hmac_sign(secret: &str, data: &[u8]) -> String {
let mut mac = HmacSha256::new_from_slice(secret.as_bytes())
.expect("HMAC accepts any key size");
mac.update(data);
hex::encode(mac.finalize().into_bytes())
}
fn hmac_verify(secret: &str, data: &[u8], signature: &str) -> bool {
let expected = hmac_sign(secret, data);
subtle::ConstantTimeEq::ct_eq(expected.as_bytes(), signature.as_bytes()).into()
}Constant-time comparison (subtle::ConstantTimeEq) prevents
timing side-channel attacks.
Initiator (client):
- Generate a random UUID nonce.
- Compute
auth_data = nonce + node_id. - Compute
auth_hmac = hmac_sign(shared_secret, auth_data). - Send
Handshake { node_id, node_name, protocol_version, agents, nonce, auth_hmac }.
Responder (server):
- Receive the
Handshakemessage. - Verify the incoming HMAC:
hmac_verify(shared_secret, nonce + node_id, auth_hmac). - If verification fails, return error code 403.
- Generate a new UUID nonce for the ack.
- Compute
ack_auth_data = ack_nonce + self.node_id. - Compute
ack_hmac = hmac_sign(shared_secret, ack_auth_data). - Send
HandshakeAck { node_id, node_name, protocol_version, agents, nonce: ack_nonce, auth_hmac: ack_hmac }.
Initiator (verification):
- Receive
HandshakeAck. - Verify:
hmac_verify(shared_secret, ack_nonce + node_id, ack_hmac). - If verification fails, return
WireError::HandshakeFailed.
| Property | How It Is Achieved |
|---|---|
| Mutual authentication | Both sides prove knowledge of the shared secret |
| Replay protection | Random UUID nonces per handshake |
| Timing-attack resistance | subtle::ConstantTimeEq for HMAC comparison |
| Mandatory secret | OFP refuses to start with an empty shared_secret |
| Message size limit | MAX_MESSAGE_SIZE = 16 MB prevents memory DoS |
| Protocol version check | PROTOCOL_VERSION mismatch returns WireError::VersionMismatch |
Source: openfang-api/src/middleware.rs
The security_headers middleware is applied to all API responses:
pub async fn security_headers(request: Request<Body>, next: Next) -> Response<Body> {
let mut response = next.run(request).await;
let headers = response.headers_mut();
headers.insert("x-content-type-options", "nosniff".parse().unwrap());
headers.insert("x-frame-options", "DENY".parse().unwrap());
headers.insert("x-xss-protection", "1; mode=block".parse().unwrap());
headers.insert("content-security-policy", /* CSP policy */);
headers.insert("referrer-policy", "strict-origin-when-cross-origin".parse().unwrap());
headers.insert("cache-control", "no-store, no-cache, must-revalidate".parse().unwrap());
response
}| Header | Value | Protects Against |
|---|---|---|
X-Content-Type-Options |
nosniff |
MIME type sniffing attacks |
X-Frame-Options |
DENY |
Clickjacking via iframes |
X-XSS-Protection |
1; mode=block |
Reflected XSS (legacy browsers) |
Content-Security-Policy |
See below | XSS, code injection, data exfiltration |
Referrer-Policy |
strict-origin-when-cross-origin |
Referrer leakage |
Cache-Control |
no-store, no-cache, must-revalidate |
Sensitive data caching |
| Directive | Value | Purpose |
|---|---|---|
default-src |
'self' |
Deny all external resources by default |
script-src |
'self' 'unsafe-inline' 'unsafe-eval' cdn.jsdelivr.net |
Allow scripts from self and CDN |
style-src |
'self' 'unsafe-inline' cdn.jsdelivr.net fonts.googleapis.com |
Allow styles from self, CDN, Google Fonts |
img-src |
'self' data: |
Allow images from self and data URIs |
connect-src |
'self' ws: wss: |
Allow WebSocket connections |
font-src |
'self' cdn.jsdelivr.net fonts.gstatic.com |
Allow fonts from CDN |
object-src |
'none' |
Block all plugins (Flash, Java, etc.) |
base-uri |
'self' |
Prevent base tag hijacking |
form-action |
'self' |
Restrict form submission targets |
Source: openfang-api/src/rate_limiter.rs
OpenFang uses the Generic Cell Rate Algorithm (GCRA) for cost-aware API
rate limiting via the governor crate.
GCRA is a leaky-bucket variant that tracks a single "virtual scheduling time" (TAT -- Theoretical Arrival Time) per key. Each request consumes a number of tokens proportional to its cost. The bucket refills at a constant rate.
Budget: 500 tokens per minute per IP address.
pub fn create_rate_limiter() -> Arc<KeyedRateLimiter> {
Arc::new(RateLimiter::keyed(Quota::per_minute(NonZeroU32::new(500).unwrap())))
}Each API operation has a configurable token cost:
pub fn operation_cost(method: &str, path: &str) -> NonZeroU32 {
match (method, path) {
(_, "/api/health") => 1,
("GET", "/api/status") => 1,
("GET", "/api/version") => 1,
("GET", "/api/tools") => 1,
("GET", "/api/agents") => 2,
("GET", "/api/skills") => 2,
("GET", "/api/peers") => 2,
("GET", "/api/config") => 2,
("GET", "/api/usage") => 3,
("GET", p) if p.starts_with("/api/audit") => 5,
("GET", p) if p.starts_with("/api/marketplace")=> 10,
("POST", "/api/agents") => 50,
("POST", p) if p.contains("/message") => 30,
("POST", p) if p.contains("/run") => 100,
("POST", "/api/skills/install") => 50,
("POST", "/api/skills/uninstall") => 10,
("POST", "/api/migrate") => 100,
("PUT", p) if p.contains("/update") => 10,
_ => 5,
}
}The cost hierarchy is intentional: read-only health checks cost 1 token while expensive operations like workflow runs cost 100, meaning a client can perform 500 health checks per minute but only 5 workflow runs.
pub async fn gcra_rate_limit(
State(limiter): State<Arc<KeyedRateLimiter>>,
request: Request<Body>,
next: Next,
) -> Response<Body> {
let ip = /* extract from ConnectInfo, default 127.0.0.1 */;
let cost = operation_cost(&method, &path);
if limiter.check_key_n(&ip, cost).is_err() {
tracing::warn!(ip, cost, path, "GCRA rate limit exceeded");
return Response::builder()
.status(StatusCode::TOO_MANY_REQUESTS)
.header("retry-after", "60")
.body(/* JSON error */)
.unwrap_or_default();
}
next.run(request).await
}pub type KeyedRateLimiter = RateLimiter<IpAddr, DashMapStateStore<IpAddr>, DefaultClock>;The DashMapStateStore provides concurrent per-IP state with automatic stale
entry cleanup.
Source: openfang-runtime/src/host_functions.rs
Two functions provide defense-in-depth against directory traversal.
Used for fs_read and fs_list operations where the target file must exist:
fn safe_resolve_path(path: &str) -> Result<std::path::PathBuf, serde_json::Value> {
let p = Path::new(path);
// Phase 1: Reject any path with ".." components
for component in p.components() {
if matches!(component, Component::ParentDir) {
return Err(json!({"error": "Path traversal denied: '..' components forbidden"}));
}
}
// Phase 2: Canonicalize to resolve symlinks and normalize
std::fs::canonicalize(p)
.map_err(|e| json!({"error": format!("Cannot resolve path: {e}")}))
}Used for fs_write operations where the target file may not exist yet:
fn safe_resolve_parent(path: &str) -> Result<std::path::PathBuf, serde_json::Value> {
let p = Path::new(path);
// Phase 1: Reject ".." in any component
for component in p.components() {
if matches!(component, Component::ParentDir) {
return Err(json!({"error": "Path traversal denied: '..' components forbidden"}));
}
}
// Phase 2: Canonicalize the parent directory
let parent = p.parent().filter(|par| !par.as_os_str().is_empty())
.ok_or_else(|| json!({"error": "Invalid path: no parent directory"}))?;
let canonical_parent = std::fs::canonicalize(parent)?;
// Phase 3: Belt-and-suspenders check on filename
let file_name = p.file_name()
.ok_or_else(|| json!({"error": "Invalid path: no file name"}))?;
if file_name.to_string_lossy().contains("..") {
return Err(json!({"error": "Path traversal denied in file name"}));
}
Ok(canonical_parent.join(file_name))
}- Capability check runs first with the raw path.
- Path traversal check runs second.
- Operation runs only if both pass.
This ordering ensures that even if a capability is misconfigured with a broad
pattern like "*", path traversal is still blocked.
Source: openfang-runtime/src/subprocess_sandbox.rs
When the runtime spawns child processes (e.g., for the shell tool or skill execution), the inherited environment must be stripped to prevent accidental leakage of secrets.
pub fn sandbox_command(cmd: &mut tokio::process::Command, allowed_env_vars: &[String]) {
cmd.env_clear(); // Remove ALL inherited env vars
// Re-add platform-independent safe vars
for var in SAFE_ENV_VARS {
if let Ok(val) = std::env::var(var) {
cmd.env(var, val);
}
}
// Re-add Windows-specific safe vars (on Windows)
#[cfg(windows)]
for var in SAFE_ENV_VARS_WINDOWS { /* ... */ }
// Re-add caller-specified allowed vars
for var in allowed_env_vars { /* ... */ }
}All platforms:
pub const SAFE_ENV_VARS: &[&str] = &[
"PATH", "HOME", "TMPDIR", "TMP", "TEMP", "LANG", "LC_ALL", "TERM",
];Windows-only:
pub const SAFE_ENV_VARS_WINDOWS: &[&str] = &[
"USERPROFILE", "SYSTEMROOT", "APPDATA", "LOCALAPPDATA",
"COMSPEC", "WINDIR", "PATHEXT",
];Variables not in these lists and not in allowed_env_vars are never
passed to the child process. This means OPENAI_API_KEY, GEMINI_API_KEY,
database credentials, and all other secrets are stripped.
pub fn validate_executable_path(path: &str) -> Result<(), String> {
let p = Path::new(path);
for component in p.components() {
if let std::path::Component::ParentDir = component {
return Err(format!(
"executable path '{}' contains '..' component which is not allowed",
path
));
}
}
Ok(())
}This prevents an agent from escaping its working directory via crafted paths
like ../../bin/dangerous.
The host_shell_exec function uses Command::new(command).args(&args) which
does not invoke a shell. Each argument is passed directly to the
process, preventing shell injection via metacharacters like ;, |, &&.
Source: openfang-skills/src/verify.rs
The SkillVerifier provides two scanning functions: security_scan() for
skill manifests and scan_prompt_content() for skill prompt text (SKILL.md
body).
SkillVerifier::security_scan(manifest) inspects a skill's declared
requirements:
| Check | Severity | Trigger |
|---|---|---|
| Node.js runtime | Warning | runtime_type == SkillRuntime::Node |
| Shell execution capability | Critical | Capability contains shellexec or shell_exec |
| Unrestricted network | Warning | Capability contains netconnect(*) |
| Shell tool | Critical | Tool is shell_exec or bash |
| Filesystem write tool | Warning | Tool is file_write or file_delete |
| Too many tools | Info | More than 10 tools required |
SkillVerifier::scan_prompt_content(content) detects common attack patterns
in skill prompt text:
Critical -- Prompt override attempts:
"ignore previous instructions", "ignore all previous",
"disregard previous", "forget your instructions",
"you are now", "new instructions:", "system prompt override",
"ignore the above", "do not follow", "override system"
Warning -- Data exfiltration patterns:
"send to http", "send to https", "post to http", "post to https",
"exfiltrate", "forward all", "send all data",
"base64 encode and send", "upload to"
Warning -- Shell command references:
"rm -rf", "chmod ", "sudo "
Info -- Excessive length:
Content over 50,000 bytes triggers an info-level warning about potential LLM performance degradation.
pub fn verify_checksum(data: &[u8], expected_sha256: &str) -> bool {
let actual = Self::sha256_hex(data);
actual == expected_sha256.to_lowercase()
}Skills installed from ClawHub have their content verified against a known SHA256 hash to detect tampering during download.
pub struct SkillWarning {
pub severity: WarningSeverity, // Info, Warning, Critical
pub message: String,
}Source: openfang-runtime/src/loop_guard.rs
The LoopGuard tracks tool calls within a single agent loop execution to
detect when the agent is stuck calling the same tool repeatedly.
pub struct LoopGuardConfig {
pub warn_threshold: u32, // Default: 3
pub block_threshold: u32, // Default: 5
pub global_circuit_breaker: u32, // Default: 30
}- For each tool call, compute SHA-256 of
tool_name + "|" + serialized_params. - Increment the count for that hash in a
HashMap<String, u32>. - Increment
total_calls. - Return a graduated verdict:
pub fn check(&mut self, tool_name: &str, params: &serde_json::Value) -> LoopGuardVerdict {
self.total_calls += 1;
// Global circuit breaker
if self.total_calls > self.config.global_circuit_breaker {
return LoopGuardVerdict::CircuitBreak(/* ... */);
}
let hash = Self::compute_hash(tool_name, params);
let count = self.call_counts.entry(hash).or_insert(0);
*count += 1;
if *count >= self.config.block_threshold {
LoopGuardVerdict::Block(/* ... */)
} else if *count >= self.config.warn_threshold {
LoopGuardVerdict::Warn(/* ... */)
} else {
LoopGuardVerdict::Allow
}
}| Verdict | Meaning | Action |
|---|---|---|
Allow |
Normal operation | Run the tool |
Warn(msg) |
Same call repeated >= 3 times | Run, append warning to result |
Block(msg) |
Same call repeated >= 5 times | Skip execution, return error |
CircuitBreak(msg) |
> 30 total tool calls | Terminate the entire agent loop |
fn compute_hash(tool_name: &str, params: &serde_json::Value) -> String {
let mut hasher = Sha256::new();
hasher.update(tool_name.as_bytes());
hasher.update(b"|");
let params_str = serde_json::to_string(params).unwrap_or_default();
hasher.update(params_str.as_bytes());
hex::encode(hasher.finalize())
}Note: serde_json::to_string produces deterministic output (object keys are
sorted), ensuring that semantically identical parameters produce the same hash.
Calls with different parameters are tracked separately. An agent that
calls web_search with 10 different queries will not trigger the guard, but
an agent that calls web_search({"query": "test"}) 5 times will be blocked.
Source: openfang-runtime/src/session_repair.rs
Before sending message history to the LLM, this module validates and repairs common structural issues that would cause API errors.
pub fn validate_and_repair(messages: &[Message]) -> Vec<Message>Phase 1 -- Collect ToolUse IDs:
Scan all messages for ContentBlock::ToolUse { id, .. } blocks and collect
their IDs into a HashSet<String>.
Phase 2 -- Filter orphans and empties:
- Orphaned ToolResults:
ContentBlock::ToolResult { tool_use_id, .. }blocks wheretool_use_idis not in the ToolUse ID set are dropped. - Empty messages: Messages with empty text or no content blocks are dropped.
Phase 3 -- Merge consecutive same-role messages:
The Anthropic API requires strict role alternation (user, assistant, user, assistant...). If two consecutive messages have the same role, they are merged into a single message with combined content blocks.
| Issue | Cause | Effect Without Repair |
|---|---|---|
| Orphaned ToolResult | Compaction or truncation removed the ToolUse | API error: "tool_use_id not found" |
| Empty messages | Cancelled generation, empty user submission | API error: empty content |
| Consecutive same-role | Manual history editing, session repair itself | API error: role alternation violation |
When merging consecutive same-role messages, both are converted to block format and concatenated:
fn merge_content(dst: &mut MessageContent, src: MessageContent) {
let dst_blocks = content_to_blocks(std::mem::replace(dst, MessageContent::Text(String::new())));
let src_blocks = content_to_blocks(src);
let mut combined = dst_blocks;
combined.extend(src_blocks);
*dst = MessageContent::Blocks(combined);
}Source: openfang-api/src/routes.rs
OpenFang provides two health endpoints with different information levels.
No authentication required. Returns only liveness information:
{
"status": "ok",
"version": "0.1.0"
}This endpoint does not expose agent count, database details, configuration warnings, uptime, or any internal system information. It is suitable for load balancer health checks.
Requires authentication. Returns full diagnostics:
{
"status": "ok",
"version": "0.1.0",
"uptime_seconds": 3600,
"panic_count": 0,
"restart_count": 2,
"agent_count": 15,
"database": "connected",
"config_warnings": []
}When no API key is configured, the auth middleware restricts all
non-health endpoints to loopback addresses only:
if api_key.is_empty() {
let is_loopback = request.extensions()
.get::<ConnectInfo<SocketAddr>>()
.map(|ci| ci.0.ip().is_loopback())
.unwrap_or(false);
if !is_loopback {
return Response::builder()
.status(StatusCode::FORBIDDEN)
.body(/* "No API key configured. Remote access denied." */)
...;
}
}# API Authentication
api_key = "your-secret-api-key" # Empty = localhost-only mode
# OFP Wire Protocol
[network]
shared_secret = "your-pre-shared-key" # Required for OFP
# WASM Sandbox
[sandbox]
fuel_limit = 1000000 # CPU instruction budget per execution
timeout_secs = 30 # Wall-clock timeout per execution
max_memory_bytes = 16777216 # 16 MB max WASM memory
# Rate Limiting
# 500 tokens/minute/IP (not currently configurable via config.toml)
# Web Search SSRF Protection
[web]
# SSRF protection is always on and cannot be disabled| Variable | Used By |
|---|---|
OPENAI_API_KEY |
OpenAI-compat driver |
ANTHROPIC_API_KEY |
Anthropic driver |
GEMINI_API_KEY or GOOGLE_API_KEY |
Gemini driver |
DEEPSEEK_API_KEY |
DeepSeek provider |
GROQ_API_KEY |
Groq provider |
BRAVE_API_KEY |
Brave web search |
TAVILY_API_KEY |
Tavily web search |
PERPLEXITY_API_KEY |
Perplexity web search |
All environment variable API keys are wrapped in Zeroizing<String> when
loaded into driver structs.
Capabilities are declared in the agent's TOML manifest:
[agent]
name = "my-agent"
[[capabilities]]
type = "FileRead"
value = "/data/*"
[[capabilities]]
type = "NetConnect"
value = "*.openai.com:443"
[[capabilities]]
type = "ToolInvoke"
value = "web_search"
[[capabilities]]
type = "LlmMaxTokens"
value = 4096The default LoopGuardConfig values are:
| Parameter | Default | Description |
|---|---|---|
warn_threshold |
3 | Identical calls before warning |
block_threshold |
5 | Identical calls before blocking |
global_circuit_breaker |
30 | Total calls before circuit break |
To pass specific environment variables to subprocesses:
sandbox_command(&mut cmd, &["MY_CUSTOM_VAR".to_string()]);Only variables explicitly listed in allowed_env_vars (plus the safe
defaults) will be inherited.
| Crate | Purpose |
|---|---|
sha2 |
SHA-256 hashing (audit trail, loop guard, SSRF, checksums) |
hmac |
HMAC-SHA256 for OFP authentication |
hex |
Hex encoding/decoding of hashes and signatures |
subtle |
Constant-time comparison (ConstantTimeEq) for HMAC verification |
ed25519-dalek |
Ed25519 signing/verification for manifest signing |
rand |
Cryptographic RNG for key generation (OsRng) |
zeroize |
Zeroizing<T> wrapper for automatic secret memory wiping |
governor |
GCRA rate limiting algorithm |
wasmtime |
WASM sandbox with fuel + epoch metering |
uuid |
Nonce generation for OFP handshakes |
chrono |
ISO-8601 timestamps for audit entries |
reqwest |
HTTP client (used inside SSRF-protected host_net_fetch) |
- sha2/hmac: Part of the RustCrypto project, audited, widely used in production Rust.
- ed25519-dalek: De facto standard Ed25519 library in Rust, extensively audited.
- subtle: Provides constant-time operations to prevent timing side-channels.
- zeroize: Official RustCrypto approach to zeroing secrets; integrates with
Drop. - governor: Battle-tested GCRA implementation with
DashMap-backed concurrent state.
| Threat | Mitigated By |
|---|---|
| Agent requests unauthorized file access | Capability-based security (Section 2) |
| Agent spawns child with elevated privileges | Capability inheritance validation (Section 2.4) |
| WASM skill runs infinite loop | Dual metering: fuel + epoch (Section 3) |
| Attacker tampers with audit log | Merkle hash chain (Section 4) |
| Prompt injection via external data | Taint tracking (Section 5) |
| Data exfiltration via LLM | Taint sinks block Secret/PII to net_fetch (Section 5.3) |
| Tampered agent manifest | Ed25519 signing (Section 6) |
| SSRF to cloud metadata | Private IP + hostname blocking + DNS check (Section 7) |
| API key recovery from memory dump | Zeroizing (Section 8) |
| Unauthorized peer-to-peer connections | HMAC-SHA256 mutual auth (Section 9) |
| XSS / clickjacking on API | Security headers (Section 10) |
| API brute force / DoS | GCRA rate limiter (Section 11) |
Path traversal via ../ |
safe_resolve_path / safe_resolve_parent (Section 12) |
| Secret leakage to child processes | env_clear() + allowlist (Section 13) |
| Malicious skills from ClawHub | Prompt injection scanner + SHA256 checksum (Section 14) |
| Agent stuck in tool loop | LoopGuard with graduated response (Section 15) |
| Corrupted LLM session history | Session repair (Section 16) |
| Information leakage from health endpoint | Redacted public endpoint (Section 17) |
| Timing attacks on HMAC verification | subtle::ConstantTimeEq (Section 9.2) |
| Shell injection via metacharacters | Command::new (no shell) + env_clear (Section 13.4) |
| DNS rebinding for SSRF bypass | Resolved IP check, not hostname check (Section 7.3) |