Loading...
Skip to main content

CVE-2026-55448 — Vulnerability in mise — CVE Details

MEDIUM

Mise's local credential_command executes untrusted config

Published June 23, 2026Updated June 23, 2026Source: osv

Summary

### Summary `mise` loads `github.credential_command` from local project config before any trust decision, then executes that value with `sh -c` when resolving a GitHub token. An attacker who can place a `.mise.toml` in a repository can execute arbitrary shell commands when the victim runs a GitHub-related mise command and no higher-priority GitHub token environment variable is set. The current command-execution path is `github.credential_command`. I confirmed in Docker that the setting is exploitable on `v2026.3.15` and `v2026.3.17`, while `v2026.3.14` rejects it as an unknown field. This report does not depend on the separate trust-bypass issue because the sink is reached directly from `[settings.github]`. ### Details The vulnerable load order is: 1. [`Settings::try_get()`](https://github.com/jdx/mise/blob/37997e70cd2216d1a86726fba0c8c09c3986ad06/src/config/settings.rs#L254-L283) preloads settings from local config files. 2. [`parse_settings_file()`](https://github.com/jdx/mise/blob/37997e70cd2216d1a86726fba0c8c09c3986ad06/src/config/settings.rs#L505-L510) returns `settings_file.settings` without checking whether the local file is trusted. 3. [`resolve_token()`](https://github.com/jdx/mise/blob/37997e70cd2216d1a86726fba0c8c09c3986ad06/src/github.rs#L344-L390) checks `settings.github.credential_command` after the token env vars and before file-based sources. 4. [`get_credential_command_token()`](https://github.com/jdx/mise/blob/37997e70cd2216d1a86726fba0c8c09c3986ad06/src/github.rs#L558-L599) executes the value with `sh -c`. The main command-execution path is: ```rust let result = std::process::Command::new("sh") .arg("-c") .arg(cmd) .arg("mise-credential-helper") .arg(host) .output() ``` If a local project file sets: ```toml [settings.github] credential_command = "echo credential_command_rce > /tmp/mise-proof.txt; echo ghp_fake_token" ``` then `resolve_token()` will reach `get_credential_command_token()` whenever higher-priority GitHub token environment variables are unset. `credential_command` is a documented custom credential source for mise, but it is also accepted from a local project `.mise.toml`, which lets an untrusted repository supply a shell command for mise to execute. ### PoC Test environment: - Docker - `linux-arm64` - `mise v2026.3.17` Negative control: ```bash export GITHUB_TOKEN=env_token mise github token --unmask ``` Observed: ```text github.com: env_token (source: GITHUB_TOKEN) /tmp/mise-proof.txt => missing ``` Primary exploit: ```toml [settings.github] credential_command = "echo credential_command_rce > /tmp/mise-proof.txt; echo ghp_fake_token" ``` Run: ```bash unset GITHUB_TOKEN GITHUB_API_TOKEN MISE_GITHUB_TOKEN MISE_GITHUB_ENTERPRISE_TOKEN mise github token --unmask ``` Observed: ```text github.com: ghp_fake_token (source: credential_command) ``` And the side effect file is created: ```text /tmp/mise-proof.txt => credential_command_rce ``` Related version check: - `v2026.3.14`: `credential_command` is rejected as an unknown field - `v2026.3.15`: the same PoC executes and returns `source: credential_command` ### Impact An attacker who can place a `.mise.toml` in a repository can execute arbitrary shell commands as the victim user when the victim runs a mise command that resolves a GitHub token from local settings. Demonstrated impact: - arbitrary command execution as the victim user - no trust prompt - no need for `[env]`, `[hooks]`, tasks, or templates Important limitation: - if a higher-priority GitHub token environment variable is already set, the `credential_command` path is not reached ### Suggested Fix Do not honor `github.credential_command` from non-global project config files. For example, inside `parse_settings_file()`: ```rust pub fn parse_settings_file(path: &Path) -> Result<SettingsPartial> { let raw = file::read_to_string(path)?; let settings_file: SettingsFile = toml::from_str(&raw)?; let mut settings = settings_file.settings; if !config::is_global_config(path) { settings.github.credential_command = None; } Ok(settings) } ```

Advisory Timeline

  • First published June 23, 2026 (111 days ago).
  • Last revised June 23, 2026 — 0 days after initial publication.

Remediation

Upgrade to the fixed version using your package manager.

Cargo
Update mise to 2026.6.4 or later
cargo update -p mise --precise 2026.6.4

After upgrading, run your dependency scanner again to confirm the vulnerability is resolved.

Affected Packages (1)

Affects 1 ecosystem: crates.io (1).

PackageEcosystemAffectedFixed In
mise
crates.io
All versions2026.6.4

Vulnerability Classification

Common Weakness Enumeration (CWE) identifiers for this vulnerability type.

  • CWE-78
    OS Command InjectionMITRE

CVSS Score Breakdown

What the CVSS (Common Vulnerability Scoring System) 6.3 score means for each attack dimension.

Attack Vector
Local
Attack Complexity
High
Privileges Required
None
User Interaction
Required
Scope
Unchanged
Confidentiality
High
Integrity
High
Availability
None

CVSS:3.1/AV:L/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:N

Frequently Asked Questions

What is CVE-2026-55448?
Mise's local credential_command executes untrusted config This vulnerability has been assigned a severity rating of MEDIUM (CVSS score: 6.3/10).
How do I check if my project is affected by CVE-2026-55448?
CVE-2026-55448 affects mise. Use GeekWala's free vulnerability scanner to check your dependencies against CVE-2026-55448 and 200,000+ other known vulnerabilities.

Severity & Exploitability

CVSS Score
6.3

Exploitation requires specific conditions or has limited impact. Remediate within weeks.

Also Known As

GHSA-29hf-rm4x-xxph

Related CVEs

  • CVE-2026-33646
    CRITICAL

    Mise Vulnerable to Arbitrary Code Execution via Tera Templates in .tool-versions Files (Trust Bypass)

  • CVE-2026-35533
    HIGH

    Local settings bypass config trust checks

  • CVE-2026-55441
    HIGH

    Mise vulnerable to arbitrary command execution via task-include files in an untrusted, config-less repository

  • CVE-2026-102821
    MEDIUM

    Russh: Unbounded memory exhaustion via CHANNEL_OPEN flood during a client-stalled rekey

  • CVE-2026-102820
    MEDIUM

    pageant: Out-of-bounds read / oversized allocation in `pageant` MemoryMap::read via a malicious Pageant agent (Windows)

Check if you're affected

Scan your dependencies to see if this vulnerability affects your projects.

Scan Your Dependencies