Security: App keys and decrypted env vars written with world-readable permissions

从 dstack-util/src/system_setup.rs 中 AppKeys 在第 1350-1351 行附近的写入，以及第 1409-1419 行附近已解密环境变量的写入开始。了解这些文件是如何打开和写入的，然后确认新写入的机密文件仅限所有者访问，而不是所有人都可读。两个写入路径都使用限制性权限，并且现有的 Rust 检查通过后，即可视为完成。

Note: This issue documents a vulnerability that was originally reported privately as the repository security advisory GHSA-535m-2x67-prhm by @pbeza.

Root Cause

fs::write() is used with default permissions (0o644) to write AppKeys to the filesystem. The AppKeys structure contains disk_crypt_key, env_crypt_key, and k256_key — all highly sensitive cryptographic secrets. Similarly, decrypted environment variables are written to files with default world-readable permissions.

// Default fs::write() creates files with 0o644 permissions
fs::write(&app_keys_path, serde_json::to_string(&app_keys)?)?;

Attack Path

1. Any process inside the CVM can read /path/to/app-keys.json
2. Attacker compromises any container workload (web vulnerability, dependency exploit, etc.)
3. Compromised process reads app-keys.json with disk_crypt_key, env_crypt_key, k256_key
4. Attacker can decrypt the disk, decrypt environment variables, and use the k256 signing key

Impact

Any compromised process inside the CVM can read all application cryptographic keys and decrypted environment variables. This includes the disk encryption key (enables offline disk reading), the env encryption key (decrypts all environment secrets), and the k256 key (enables signing as the CVM).

Suggested Fix

Set restrictive permissions before writing:

use std::os::unix::fs::OpenOptionsExt;

let mut file = std::fs::OpenOptions::new()
    .write(true)
    .create(true)
    .truncate(true)
    .mode(0o600)
    .open(&app_keys_path)?;
file.write_all(serde_json::to_string(&app_keys)?.as_bytes())?;

Consider further restricting to 0o400 (read-only by owner) after the initial write.

Note: This finding was reported automatically as part of an AI/Claude-driven internal audit by the NEAR One MPC team. It has not been manually verified by a human to confirm whether it constitutes an actual security issue.