Skip to main content

vtcode_safety/sandboxing/
child_spawn.rs

1//! Child process spawning with sandbox-aware environment handling.
2//!
3//! Implements patterns from the Codex sandbox model:
4//! - Environment variable sanitization (remove sensitive vars)
5//! - Parent death signal (PR_SET_PDEATHSIG on Linux)
6//! - Sandbox identification markers for downstream tools
7
8use hashbrown::HashMap;
9use std::path::Path;
10
11/// Environment variables that should be filtered from sandboxed processes.
12///
13/// Following the field guide: "Completely clear the environment and rebuild it
14/// with only the variables you actually want."
15pub const FILTERED_ENV_VARS: &[&str] = &[
16    // API keys and tokens
17    "OPENAI_API_KEY",
18    "ANTHROPIC_API_KEY",
19    "GEMINI_API_KEY",
20    "XAI_API_KEY",
21    "DEEPSEEK_API_KEY",
22    "META_API_KEY",
23    "MODEL_API_KEY",
24    "OPENROUTER_API_KEY",
25    "GROQ_API_KEY",
26    "MISTRAL_API_KEY",
27    "COHERE_API_KEY",
28    "AZURE_OPENAI_API_KEY",
29    "HUGGINGFACE_API_KEY",
30    "HF_TOKEN",
31    // Cloud provider credentials
32    "AWS_ACCESS_KEY_ID",
33    "AWS_SECRET_ACCESS_KEY",
34    "AWS_SESSION_TOKEN",
35    "GOOGLE_APPLICATION_CREDENTIALS",
36    "GOOGLE_CLOUD_PROJECT",
37    "AZURE_CLIENT_ID",
38    "AZURE_CLIENT_SECRET",
39    "AZURE_TENANT_ID",
40    "AZURE_SUBSCRIPTION_ID",
41    // GitHub tokens
42    "GITHUB_TOKEN",
43    "GH_TOKEN",
44    "GITHUB_PAT",
45    // NPM/Package registry tokens
46    "NPM_TOKEN",
47    "NPM_AUTH_TOKEN",
48    "CARGO_REGISTRY_TOKEN",
49    "PYPI_TOKEN",
50    // Database credentials
51    "DATABASE_URL",
52    "DB_PASSWORD",
53    "PGPASSWORD",
54    "MYSQL_PWD",
55    "REDIS_PASSWORD",
56    "MONGO_PASSWORD",
57    // SSH/GPG
58    "SSH_AUTH_SOCK",
59    "GPG_AGENT_INFO",
60    // Dynamic linker vars (security risk)
61    "LD_PRELOAD",
62    "LD_LIBRARY_PATH",
63    "LD_AUDIT",
64    "LD_DEBUG",
65    "LD_PROFILE",
66    "DYLD_INSERT_LIBRARIES",
67    "DYLD_LIBRARY_PATH",
68    "DYLD_FRAMEWORK_PATH",
69    "DYLD_FALLBACK_LIBRARY_PATH",
70    // Other sensitive vars
71    "VAULT_TOKEN",
72    "CONSUL_HTTP_TOKEN",
73    "DOCKER_AUTH_CONFIG",
74    "KUBECONFIG",
75    "KUBE_TOKEN",
76    "SLACK_TOKEN",
77    "SLACK_BOT_TOKEN",
78    "DISCORD_TOKEN",
79    "TELEGRAM_BOT_TOKEN",
80];
81
82/// Environment variables that should always be preserved.
83pub const PRESERVED_ENV_VARS: &[&str] = &[
84    // Basic shell environment
85    "PATH",
86    "HOME",
87    "USER",
88    "SHELL",
89    "TERM",
90    "LANG",
91    "LC_ALL",
92    "LC_CTYPE",
93    "TZ",
94    // XDG directories
95    "XDG_CONFIG_HOME",
96    "XDG_DATA_HOME",
97    "XDG_CACHE_HOME",
98    "XDG_RUNTIME_DIR",
99    // Editor preferences (not sensitive)
100    "EDITOR",
101    "VISUAL",
102    "PAGER",
103    // Build tool paths
104    "CARGO_HOME",
105    "RUSTUP_HOME",
106    "GOPATH",
107    "GOROOT",
108    "JAVA_HOME",
109    "PYTHON",
110    "PYTHONPATH",
111    "NODE_PATH",
112    // Terminal capabilities
113    "COLORTERM",
114    "FORCE_COLOR",
115    "NO_COLOR",
116    "CLICOLOR",
117    "CLICOLOR_FORCE",
118    // Temp directories
119    "TMPDIR",
120    "TEMP",
121    "TMP",
122];
123
124/// Sandbox environment markers set for child processes.
125pub const VTCODE_SANDBOX_ACTIVE: &str = "VTCODE_SANDBOX_ACTIVE";
126pub const VTCODE_SANDBOX_NETWORK_DISABLED: &str = "VTCODE_SANDBOX_NETWORK_DISABLED";
127pub const VTCODE_SANDBOX_TYPE: &str = "VTCODE_SANDBOX_TYPE";
128pub const VTCODE_SANDBOX_WRITABLE_ROOTS: &str = "VTCODE_SANDBOX_WRITABLE_ROOTS";
129
130/// Build a sanitized environment for sandboxed child processes.
131///
132/// Implements the Codex pattern: "Completely clear the environment and rebuild it
133/// with only the variables you actually want."
134#[expect(
135    unused_results,
136    reason = "Environment construction intentionally ignores prior values while writing the sanitized snapshot."
137)]
138pub fn build_sanitized_env(
139    current_env: &HashMap<String, String>,
140    sandbox_active: bool,
141    network_disabled: bool,
142    sandbox_type: &str,
143    writable_roots: &[&Path],
144) -> HashMap<String, String> {
145    let mut sanitized = HashMap::new();
146
147    // Copy only preserved environment variables
148    for key in PRESERVED_ENV_VARS {
149        if let Some(value) = current_env.get(*key) {
150            sanitized.insert(key.to_string(), value.clone());
151        }
152    }
153
154    // Add sandbox markers so downstream tools know what's happening
155    if sandbox_active {
156        sanitized.insert(VTCODE_SANDBOX_ACTIVE.to_string(), "1".to_string());
157        sanitized.insert(VTCODE_SANDBOX_TYPE.to_string(), sandbox_type.to_string());
158
159        if network_disabled {
160            sanitized.insert(VTCODE_SANDBOX_NETWORK_DISABLED.to_string(), "1".to_string());
161        }
162
163        if !writable_roots.is_empty() {
164            let roots: Vec<String> = writable_roots.iter().map(|p| p.display().to_string()).collect();
165            sanitized.insert(VTCODE_SANDBOX_WRITABLE_ROOTS.to_string(), roots.join(":"));
166        }
167    }
168
169    sanitized
170}
171
172/// Check if an environment variable should be filtered.
173pub fn should_filter_env_var(key: &str) -> bool {
174    let key = key.to_ascii_uppercase();
175    FILTERED_ENV_VARS.contains(&key.as_str())
176        || key.starts_with("AWS_")
177        || key.starts_with("AZURE_")
178        || key.starts_with("GOOGLE_")
179        || key.starts_with("GCP_")
180        || key.starts_with("CLOUDSDK_")
181        || key.starts_with("LD_")
182        || key.starts_with("DYLD_")
183        || matches!(key.as_str(), "TOKEN" | "SECRET" | "PASSWORD" | "PASS" | "PWD" | "CREDENTIALS")
184        || key.ends_with("_TOKEN")
185        || key.ends_with("_KEY")
186        || key.ends_with("_SECRET")
187        || key.ends_with("_PASS")
188        || key.ends_with("_PWD")
189        || key.ends_with("_PASSWORD")
190        || key.ends_with("_CREDENTIALS")
191}
192
193/// Filter sensitive environment variables from an existing map.
194///
195/// Less aggressive than `build_sanitized_env` - preserves most vars but removes known sensitive ones.
196pub fn filter_sensitive_env(env: &HashMap<String, String>) -> HashMap<String, String> {
197    env.iter()
198        .filter(|(k, _)| !should_filter_env_var(k))
199        .map(|(k, v)| (k.clone(), v.clone()))
200        .collect()
201}
202
203/// Set up parent death signal on Linux.
204///
205/// "Ensures sandboxed children die if the main process gets killed -
206/// you don't want orphaned processes running around."
207///
208/// Uses SIGTERM for graceful shutdown. Includes parent PID check to avoid
209/// race condition where parent exits between fork and exec.
210#[cfg(target_os = "linux")]
211pub fn setup_parent_death_signal() -> std::io::Result<()> {
212    setup_parent_death_signal_with_check(nix::unistd::getppid())
213}
214
215/// Set up parent death signal with explicit parent PID check.
216///
217/// This variant should be used in pre_exec hooks where the parent PID
218/// is captured before spawn to avoid race conditions.
219#[cfg(target_os = "linux")]
220pub fn setup_parent_death_signal_with_check(expected_parent_pid: nix::unistd::Pid) -> std::io::Result<()> {
221    use nix::sys::prctl;
222    use nix::sys::signal::{Signal, raise};
223    use std::io::Error;
224
225    // Use SIGTERM for graceful shutdown (allows cleanup handlers to run)
226    prctl::set_pdeathsig(Some(Signal::SIGTERM))
227        .map_err(|e| Error::other(format!("prctl(PR_SET_PDEATHSIG) failed: {e}")))?;
228
229    // Re-check parent PID to catch race condition where parent exited between
230    // fork and this prctl call. If parent changed, self-terminate immediately.
231    if nix::unistd::getppid() != expected_parent_pid {
232        let _ = raise(Signal::SIGTERM);
233    }
234
235    Ok(())
236}
237
238#[cfg(not(target_os = "linux"))]
239pub fn setup_parent_death_signal() -> std::io::Result<()> {
240    Ok(())
241}
242
243#[cfg(test)]
244mod tests {
245    use super::*;
246
247    const TEST_API_KEY_VALUE: &str = "test-openai-key";
248
249    #[test]
250    fn test_should_filter_sensitive_vars() {
251        assert!(should_filter_env_var("OPENAI_API_KEY"));
252        assert!(should_filter_env_var("META_API_KEY"));
253        assert!(should_filter_env_var("AWS_SECRET_ACCESS_KEY"));
254        assert!(should_filter_env_var("GITHUB_TOKEN"));
255        assert!(should_filter_env_var("LD_PRELOAD"));
256        assert!(should_filter_env_var("DYLD_INSERT_LIBRARIES"));
257        assert!(should_filter_env_var("MY_CUSTOM_TOKEN"));
258        assert!(should_filter_env_var("token"));
259        assert!(should_filter_env_var("CloudSDK_AUTH_CREDENTIAL_FILE_OVERRIDE"));
260        assert!(should_filter_env_var("MY_CUSTOM_PASS"));
261        assert!(should_filter_env_var("MYSQL_PWD"));
262        assert!(should_filter_env_var("DATABASE_PASSWORD"));
263
264        assert!(!should_filter_env_var("PATH"));
265        assert!(!should_filter_env_var("HOME"));
266        assert!(!should_filter_env_var("TERM"));
267    }
268
269    #[test]
270    fn filters_sensitive_names_case_insensitively() {
271        assert!(should_filter_env_var("openai_api_key"));
272        assert!(should_filter_env_var("Aws_SECRET_ACCESS_KEY"));
273        assert!(should_filter_env_var("custom_access_token"));
274        assert!(should_filter_env_var("dyld_insert_libraries"));
275        assert!(!should_filter_env_var("PROJECT_NAME"));
276    }
277
278    #[test]
279    fn test_build_sanitized_env() {
280        let mut current = HashMap::new();
281        drop(current.insert("PATH".to_string(), "/usr/bin".to_string()));
282        drop(current.insert("HOME".to_string(), "/home/user".to_string()));
283        drop(current.insert("OPENAI_API_KEY".to_string(), TEST_API_KEY_VALUE.to_string()));
284        drop(current.insert("RANDOM_VAR".to_string(), "value".to_string()));
285
286        let sanitized = build_sanitized_env(&current, true, true, "MacosSeatbelt", &[]);
287
288        // PATH and HOME should be preserved
289        assert_eq!(sanitized.get("PATH"), Some(&"/usr/bin".to_string()));
290        assert_eq!(sanitized.get("HOME"), Some(&"/home/user".to_string()));
291
292        // API key should NOT be present (not in preserved list)
293        assert!(!sanitized.contains_key("OPENAI_API_KEY"));
294
295        // Random var should NOT be present (not in preserved list)
296        assert!(!sanitized.contains_key("RANDOM_VAR"));
297
298        // Sandbox markers should be set
299        assert_eq!(sanitized.get(VTCODE_SANDBOX_ACTIVE), Some(&"1".to_string()));
300        assert_eq!(sanitized.get(VTCODE_SANDBOX_NETWORK_DISABLED), Some(&"1".to_string()));
301        assert_eq!(sanitized.get(VTCODE_SANDBOX_TYPE), Some(&"MacosSeatbelt".to_string()));
302    }
303
304    #[test]
305    fn test_filter_sensitive_env() {
306        let mut env = HashMap::new();
307        drop(env.insert("PATH".to_string(), "/usr/bin".to_string()));
308        drop(env.insert("OPENAI_API_KEY".to_string(), TEST_API_KEY_VALUE.to_string()));
309        drop(env.insert("MY_VAR".to_string(), "value".to_string()));
310        drop(env.insert("AWS_ACCESS_KEY_ID".to_string(), "AKIA...".to_string()));
311        drop(env.insert("CUSTOM_PASS".to_string(), "let-me-in".to_string()));
312        drop(env.insert("SERVICE_PWD".to_string(), "super-secret".to_string()));
313
314        let filtered = filter_sensitive_env(&env);
315
316        assert!(filtered.contains_key("PATH"));
317        assert!(filtered.contains_key("MY_VAR"));
318        assert!(!filtered.contains_key("OPENAI_API_KEY"));
319        assert!(!filtered.contains_key("AWS_ACCESS_KEY_ID"));
320        assert!(!filtered.contains_key("CUSTOM_PASS"));
321        assert!(!filtered.contains_key("SERVICE_PWD"));
322    }
323
324    #[test]
325    fn filters_adversarial_credential_and_loader_names() {
326        let mut env = HashMap::new();
327        for (name, value) in [
328            ("MODEL_API_KEY", "model-secret"),
329            ("META_API_KEY", "meta-secret"),
330            ("INSTRUCTION_SECRET", "prompt-secret"),
331            ("AWS_PROFILE", "production"),
332            ("LD_AUDIT", "audit.so"),
333            ("DYLD_LIBRARY_PATH", "/tmp/injected"),
334            ("SAFE_PROJECT_NAME", "vtcode"),
335        ] {
336            drop(env.insert(name.to_string(), value.to_string()));
337        }
338
339        let filtered = filter_sensitive_env(&env);
340        let rebuilt = build_sanitized_env(&env, true, false, "test", &[]);
341
342        for sensitive in [
343            "MODEL_API_KEY",
344            "META_API_KEY",
345            "INSTRUCTION_SECRET",
346            "AWS_PROFILE",
347            "LD_AUDIT",
348            "DYLD_LIBRARY_PATH",
349        ] {
350            assert!(!filtered.contains_key(sensitive), "sensitive variable survived filtering: {sensitive}");
351            assert!(!rebuilt.contains_key(sensitive), "sensitive variable survived rebuilding: {sensitive}");
352        }
353        assert_eq!(filtered.get("SAFE_PROJECT_NAME"), Some(&"vtcode".to_string()));
354    }
355}