Gemini CLI
Google · CLI agent · homepage · google-gemini/gemini-cli
Minimum safe version is set by CVE-2026-12537 (Gemini CLI: Remote Code Execution via workspace trust and tool allowlisting bypasses, fixed in 0.39.1): the highest fix among Gemini CLI's medium-or-higher advisories. Only advisories for Gemini CLI itself are counted; other components and ecosystem projects are listed separately below.
Advisories · Data exposure · Hardening
Advisories
| ID | Title | Severity | Fixed in | Published |
|---|---|---|---|---|
| CVE-2026-12537 | Gemini CLI: Remote Code Execution via workspace trust and tool allowlisting bypasses · sets minimum safe version | Critical | 0.39.1 | 2026-04-24 |
Related products and ecosystem projects
Advisories for other Google components or third-party projects that build on Gemini CLI. Not counted above and not used for the minimum safe version.
| ID | Component | Title | Severity | Fixed in | Published |
|---|---|---|---|---|---|
| CVE-2026-117560 | ATXP Gemini CLI extension | ATXP Gemini CLI extension through 1.0.0 contains an OS command injection vulnerability in runAtxp() that allows attackers to execute shell c… | High | — | 2026-10-27 |
| CVE-2026-54785 | gemini-bridge | gemini-bridge is a lightweight MCP server bridging AI agents to Google's Gemini AI via the official CLI. From 1.0.0 until 1.3.1, consult_gem… | Medium | — | 2026-07-31 |
Hardening checklist
- Keep auto-update enabled and verify the installed version against the minimum safe version.
- Do not open untrusted repositories with agent mode or auto-approve enabled.
- Restrict MCP servers to an allowlist; review each server's permissions and publisher.
- Never place secrets in project files the agent can read; use a secrets manager or env injection.
- Run with sandboxing enabled when working on untrusted repositories.
Data exposure
What Gemini CLI keeps on a developer machine, what it sends, and the settings that reduce exposure. Local transcripts can contain anything pasted into a session, including secrets: include these paths in secret scanning and offboarding.
Stored locally
| What | Where |
|---|---|
| Full session history: prompts, responses, all tool inputs and outputs, and token stats. | ~/.gemini/tmp/<project_hash>/chats/ |
| Checkpoints (off by default): a Git snapshot of project files plus the conversation and pending tool call. | ~/.gemini/history/<project_hash> (shadow Git repo) and ~/.gemini/tmp/<project_hash>/checkpoints |
| Cached Google sign-in (OAuth) credentials. | ~/.gemini/oauth_creds.json |
API keys, if you place them in a .env file as the docs suggest. | ~/.gemini/.env or .gemini/.env in the project |
Sent to the vendor
Prompts and context go to the Google service behind your sign-in method (Gemini Code Assist, Gemini Developer API or Vertex AI). Usage statistics are sent by default; Google says they exclude prompt and response content, file contents, PII and API keys.
Retention and training
Terms depend on the sign-in method. Gemini API unpaid tier: content may be used to improve Google products and read by human reviewers. Paid tier: not used for improvement, logged for a limited period for abuse detection. Code Assist Standard/Enterprise and Vertex AI follow Google Cloud terms. Consumer "Login with Google" access ended June 18, 2026.
Settings that reduce exposure
general.sessionRetention.maxAge(default"30d")- Deletes local sessions older than the set age;
maxCountalso caps the number kept. general.checkpointing.enabled(defaultfalse)- When enabled, stores project snapshots and conversation copies under
~/.gemini/. privacy.usageStatisticsEnabled: false- Opts out of usage statistics sent to Google.
telemetry.logPrompts: false- Keeps user prompt content out of telemetry logs when telemetry is enabled.
For administrators
- System settings file overrides user settings:
/etc/gemini-cli/settings.json(Linux),C:\ProgramData\gemini-cli\settings.json(Windows),/Library/Application Support/GeminiCli/settings.json(macOS). - Enterprise Admin Controls (Management Console) cannot be overridden locally: Strict Mode blocks YOLO mode, and extensions and MCP servers can be disabled or allowlisted.
security.auth.enforcedTypeforces one authentication method for all users.- A proxy adding the
X-GoogApps-Allowed-Domainsheader blocks sign-in with non-corporate Google accounts.
Sources: Gemini CLI docs: Session management · Gemini CLI docs: Checkpointing · Gemini CLI docs: Configuration reference (usage statistics) · Gemini CLI docs: Authentication · Gemini CLI docs: Gemini CLI for the enterprise · Gemini CLI docs: Enterprise Admin Controls · Gemini CLI docs: Terms of Service and Privacy Notice · Gemini API Additional Terms of Service · Google: Gemini Code Assist for individuals deprecation · BoostSecurity Bagel: AI CLI probe (Gemini credential and chat paths) · reviewed 2026-11-09
Rolling these tools out? See the playbook Secure AI coding tools in your organisation.
Lock it down: AI hardening kit (managed settings and MCP allowlist generators, free, in your browser).
Check your version
Via the public API:
curl -s "https://threatdeck.net/api/v1/tools/gemini-cli/check?version=<your-version>" Or ask your agent through the ThreatDeck MCP server (https://threatdeck.net/mcp):
check_ai_tool_version { "slug": "gemini-cli", "version": "<your-version>" } Updated 2026-11-22